# Index — `~/code/`

_Single navigable catalog of every load-bearing markdown file. Maintained by Claude. Updated when files are added, removed, or significantly restructured. For chronological "what changed when," see `log.md`._

_Last full pass: 2026-05-17._

---

- `docs/rules/focus-board.md` — Focus board rules, verbatim history, IDs (moved from CLAUDE.md + memory 2026-09-03)
- `docs/rules/supabase-security.md` — Supabase security rules + the view-exposure incidents (moved from CLAUDE.md 2026-09-03)
- `docs/rules/surface-status.md` — **who can see each surface TODAY** (nobody · staff · customers) + the applied/built/shipped vocabulary; read before choosing a migration lane. Born 2026-09-17: Passport was treated as live all week; it has no audience yet.
- `docs/rules/outside-voices.md` — **Boris/Karpathy tips ledger**: each outside tip checked against what we already do (have · gap · trial · pass · landed). Screenshot intake. Born 2026-10-18.
- `docs/rules/glossary.md` — **the words that vacillate between terminals**: applied · built · shipped (+ audience); "in production" is banned (Kyle means *being made*, terminals hear *live*). Linted by `hygiene-lint.sh` §1f.
- `ThickAsThieves/DROPS_ARCHITECTURE.md` — drops subsystem map: what talks to what, writers/readers/push rails/schedules, known drift (2026-09-03)
- [ThickAsThieves/docs/CORKY.md](ThickAsThieves/docs/CORKY.md) — Corky, the somm in the TaT app: design of record (shapes not a prompt box, per-user pours, PCC-first prose, Haiku on the credit for Android, scored on the till, cellar-tracker import as the named upgrade). **Merged + applied 2026-10-09 with `enabled=false`; the four Codex round-2 pre-ship fixes landed 2026-10-12 (§Pre-ship fixes; package `2026-10-12-0002-corky-preship`); same day Kyle flipped it on for his builds and ruled the open question box in: `ask` shape + character + asks catalogue built (§Built 2026-10-12; package `2026-10-13-0001-corky-ask`, `scripts/corky-asks.mjs`).**
- `docs/reviews/2026-09-14-fable-burn-diagnosis.md` — MEASURED: where a week's Fable went (76% into two migration-package builds), who wrote the ceremony, the three-lane ruling + retire list; re-runnable method
- `docs/reviews/2026-09-13-ceremony-recalibration.md` — the brief that asked whether migration ceremony is miscalibrated (answered by the 09-14 diagnosis)
- `docs/proposals/2026-09-03-claude-md-trim.md` — proposal: trim CLAUDE.md ~45% to map + rules, narratives to rule files (Kyle ratifies)
- `db-migrations/_templates/survey.md` + `survey.sql` — Gate 0.5 survey template (hook-required for packages ≥ 2026-09-04)
- `db-migrations/_templates/transition-matrix.sql` — **NEW 2026-09-12.** Gate 3 row-diff harness for any migration that MOVES rows. A control that grades a total is blind to a swap that conserves it — twice in one week (the log rotation's conservation invariant duplicating 21 archive entries; a 111-line wine silently changing AVAs under all-green assertions). Snapshot keyed before → apply → snapshot after → FULL OUTER JOIN → read the matrix.

## Master orientation (read first, every session)

- [CLAUDE.md](CLAUDE.md) — lean session-start **cockpit**: orientation, the "which file to reference" map, money/prod warnings, credentials. Auto-loaded; **@-imports `SPINE.md`**. (Restructured into layers 2026-06-15 — routing/audit in `docs/rebuild/RECONCILIATION.md`; prior version preserved as `docs/archive/2026-09/CLAUDE.v1-backup.md`.)
- [SPINE.md](SPINE.md) — the operating **constitution**: principles, the discipline (the wiki layer), and the landmine *laws* (the expensive lessons), as `[U]`/`[U+i]`/`[T]`-tagged items. @-imported by CLAUDE.md, so always in force. The `[U]` subset is the Project-B general spine.
- [AGENTS.md](AGENTS.md) — **NEW 2026-06-20.** The bridge that gives Codex (and any non-Claude agent) our context — Codex auto-reads `AGENTS.md`, **not** `CLAUDE.md`. Routes it to read `SPINE.md`/`CLAUDE.md`/`log.md`, defines its lane (advisor + creative/image + scoped/greenfield builds in its own git worktree; Claude owns `main` / the wiki layer / money-prod / multi-file work), and bakes in the guardrails (never `--yolo`, never touch metered/prod, never edit the wiki layer → propose-only, Claude lands). Bootstrap prompt for a fresh Codex session: `docs/codex-starter-prompt.md`. Run-commands + worktree recipe live in CHEATSHEET "AI workshop"; rationale in the `reference_codex-found-compute` memory.
- [CLAUDE_ARCHIVE.md](CLAUDE_ARCHIVE.md) — completed work history. Append-only; entries dated.
- [CLAUDE_SUPABASE.md](CLAUDE_SUPABASE.md) — full Supabase schema with column-level detail.
- [db-docs/](db-docs/) — **NEW 2026-06-13.** tbls auto-generated schema docs: every public table/view (116), columns/types/FKs/indexes + a Mermaid ER diagram, introspected straight from the live DB. The *structural* truth layer that never drifts — the companion to CLAUDE_SUPABASE.md's *meaning* layer (which tables are living vs frozen, business semantics). Read-only role `tbls_ro`; DSN in `~/.zshenv`, never committed. Regenerate: `cd db-docs && tbls doc --force`. Also holds `DEBRIS.md` (auto-generated stash/backup report) and, from the 2026-07-19 census, [TABLE_CENSUS_2026-07-19.md](db-docs/TABLE_CENSUS_2026-07-19.md) + `proposed-comments-2026-07-19.sql`.
- [db-docs/TABLE_CENSUS_2026-07-19.md](db-docs/TABLE_CENSUS_2026-07-19.md) — **NEW 2026-07-19.** Liveness/ownership census of all 234 public objects: what's LIVING vs FROZEN vs UNBUILT vs droppable, who writes and reads each table. Read it before assuming an empty table is dead — 15 objects are *unbuilt*, not abandoned. Companion `proposed-comments-2026-07-19.sql` = 163 drafted `COMMENT ON` statements awaiting Kyle's ratify (applying them takes table-comment coverage 30% → 99%, and tbls then renders them into the per-table docs automatically).
- [db-migrations/README.md](db-migrations/README.md) — **NEW 2026-08-22, the factory-repair contract.** The canonical migration registry + Gate 0–7 lifecycle for every production DDL change: durable `migration.sql` (hash-matched by `~/.claude/hooks/migration-guard.mjs`, which DENIES novel DDL through `apply_migration`), per-statement object lint, rehearsal ladder (BEGIN…ROLLBACK rung 1 → Supabase branch rung 2), in-migration catalog assertions, hash-bound cross-vendor review for exposure-changing work, Kyle release, production read-back. Epoch boundary + named blind spots inside. First passage: `2026-08-22-0001-close-public-creation-defaults/` (new objects born closed to clients).
- [CLAUDE_AI_MEMORY.md](CLAUDE_AI_MEMORY.md) — preserved snapshot of claude.ai's memory (the phone/brainstorm surface of the three-Claude team). claude.ai's real memory is model-managed + capped (~30×500 chars), so this is the full version + the source Kyle re-pastes from. Dated snapshot — refresh on drift.
- [LESSONS_LEARNED.md](LESSONS_LEARNED.md) — incident post-mortems. **Never archive, never compress, never touched on cleanup days.**
- [SECRETS_LOG.md](SECRETS_LOG.md) — every historical secret exposure + rotation status. Read before rotating any credential.
- [SHOPIFY_APPS.md](SHOPIFY_APPS.md) — custom-app registry: app name ↔ client ID ↔ which repos authenticate as it. Read before touching any Shopify credential or assuming an app is dedicated (the admin `SHOPIFY_CLIENT_ID` is shared across 4+ repos).
- [SHOPIFY_API_VERSION_AUDIT.md](SHOPIFY_API_VERSION_AUDIT.md) — **NEW 2026-06-13. Handoff brief.** Cross-repo Shopify API-version audit, framed as the *silent-failure triple* (old pinned version × version-gated feature × swallowed errors). Key finding: `shopify-xero-sync` pinned `2025-01` (prod accounting layer, sunset exposure). Map → triage → bump per-repo, the auto-deploy repo last on a branch.
- [SECOND_STORE_READINESS_AUDIT.md](SECOND_STORE_READINESS_AUDIT.md) — **NEW 2026-07-19. Horizon B4 read-only audit.** Central reframe: NS ("The Thief Spirits", location `84585578651`) has been TRANSACTING since 2026-03-14 — the one-location assumptions are live defects, not future risks. F1: GG receiving routes by product type not destination → 12 split-brain SKUs (NS −118 / WW +338 phantom). F2: restock hunt has no location dimension. F3: reorder engine triggers on aggregate inventory. F4: Xero sees one store. Ends in 3 Kyle rulings. Nothing fixed by design.
- [CHEATSHEET.md](CHEATSHEET.md) — quick reference: run commands, project locations, daily ops.
- [RULINGS.md](RULINGS.md) — **NEW 2026-08-05, ⚠ PARKED/DRAFT — not doctrine, do not route rulings here yet** (the `CLAUDE.md` gate-card loop still governs; Kyle reviews the format with fresh eyes). The rulings punch list: every open decision needing Kyle, grouped by project, one line each, **with the terminal's call already pre-filled**. Born because the Focus board hit 72 open Your-Move cards, **27 ruling-shaped** — a card is the wrong container for a decision (title + Notes blob + a file path Kyle must hunt), so rulings sat for weeks and he couldn't see what they blocked. **Cards stay for linear work; decisions come here.** Kyle says "rule on TAT" → a terminal recites that group and applies the answers same session. Writing rule: a line with no pre-filled call is unfinished analysis, not a question.
- [docs/terminal-primer.md](docs/terminal-primer.md) — **NEW 2026-07-25.** Kyle-facing terminal orientation, written mid-mini-setup off a real complaint ("I never know what's included and what's just spaces"). Nine facts: reading the prompt line · **how to copy a fenced command** (the ` ```bash ` language tag is NOT part of the command — it caused a live failure) · Windows→macOS verb table · ⚠ `Ctrl+C` kills, `Cmd+C` copies · Tab/↑/`pwd`/`ls` · silence=success · invisible password typing · paste whole output, never a summary. Point Kyle here when he's at an unfamiliar box.
- [MACHINES.md](MACHINES.md) — **NEW 2026-08-31, written from the Windows PC on its first session.** The machine registry: which boxes exist (laptop · zanzibar M4 Pro · staff M2 · incoming M5 Max Studio · the Windows PC · UDM Pro Max), what's installed on each, and which one should run what. **Every row carries a `[measured]` / `[wiki]` / `[Kyle]` provenance tag** — a box can only measure itself, so only the PC section is measured and every Mac row still wants a verifying pass. Carries the PC's ⚠ gaps (no guard hooks → **Supabase MCP is live and ungated there**; the `~/.claude` harness exists in exactly one unbacked-up place; Win10 past end-of-support), the **don't-ferry-.env ruling** (→ finish the 1Password `op://` rail instead), and the honest read on its RTX 3070 — calibrated against `zanzibar/bench/VERDICT_PHASE2.md`, which already ruled local models *matching/classification-grade, not extraction*. Read before deciding which machine a job belongs on.
- [DATABASE.md](DATABASE.md) — plain-English cheat sheet of Supabase tables and how they connect.
- [ENRICHMENT_DATA_MAP.md](ENRICHMENT_DATA_MAP.md) — the "card catalog" for enrichment data: verified real `count(*)` (list_tables row counts LIE), active-wine coverage table, the **keys-not-gold** finding, where each signal lives + its provenance column, the FK/soft-key join map (**the spine has NO FK — that's the leak**), caution log, refresh queries. **Read before quoting any enrichment coverage number.** Created 2026-06-06.
- [thief-wd-rescue/CORPUS_IDENTITY_DESIGN.md](thief-wd-rescue/CORPUS_IDENTITY_DESIGN.md) — **NEW 2026-08-18. Design of record for corpus identity: the ladder (v2 normalizer → designation split-back → UPC∧containment → reorder → survivor pin → accumulative `corpus_key_aliases` → `corpus_twins` → `corpus_publish_v2` atomic re-key with ledger) + the GG enrichment cascade (importer/grape/note/image rungs, corpus pin, /gaps attach). Carries both Codex reviews (21 design findings + 11 implementation findings) with rulings. Read before touching `wine_key`, `normName`, or any corpus consumer. Migration: `thief-wd-rescue/migrations/2026-08-18-corpus-identity-aliases.sql`; normalizer: `thief-wd-rescue/scripts/lib/wine-identity.mjs` (+ copies, parity script `scripts/check-wine-identity-parity.sh`).
- [DISTRIBUTOR_COALESCE_DESIGN.md](DISTRIBUTOR_COALESCE_DESIGN.md) — the (wine, vintage) identity model + per-source coalesce rules for `distributor_products` → `distributor_wines`. Identity = wine+vintage NOT UPC; field grains (producer/wine/vintage); **shared-UPC ≠ same product inside distributor data** (SGWS reuses one UPC across 61 sakes); per-source vintage extraction; keg/quarantine rules. Engine: `distributor-scrape/src/coalesce/`. Verified by two adversarial sweeps 2026-06-09. Read before consuming `distributor_products` or building corpus landing.
- [BUILD_CHARTER.md](BUILD_CHARTER.md) — short build ethos: read first, deterministic first, no happy-path code, two-strike stop, finish the last 10%, and hard gates for money/prod/external work.
- [MAJOR_CHANGE_GATE.md](MAJOR_CHANGE_GATE.md) — pre-flight checklist for any major refactor/rewrite/repo-split: name the original pain, define what must be different, prove the plan is not rebuilding the old problem with new files.
- [buying-guardrails.md](buying-guardrails.md) — two-tier SKU cap system. Per-region floor caps, Cave/Billboard rules, vintage rollover policy.
- [THIEF_HOSPITALITY.md](THIEF_HOSPITALITY.md) — service vs hospitality, the Rose move, collect/connect dots, "unreasonable for free." Read when writing customer-facing copy, designing staff/customer UX, or sketching tasting events. Sources: NotebookLM "Hospitalian Way" notebook (Stuckey/Meyer/Guidara).
- [JUDGMENT_RUBRICS.md](JUDGMENT_RUBRICS.md) — **NEW 2026-07-16, Fable exit-artifact.** The taste layer written down: decision PROCEDURES (not outcomes) for judgment queues — wine-identity matching (producer-first, tier-words-never-bridge, decoration-does-bridge, never-bank-onto-code-rows, stale-pool re-check), confidence semantics, overturn rules, the needs_kyle escalation line. Seeded from the live 143-row corpus pre-grade sitting; extend one rubric per queue-type as sittings happen.
- [YEAR_END_COUNT_SPEC.md](project-fawkes/docs/YEAR_END_COUNT_SPEC.md) — **NEW 2026-11-04, reviewed spec.** Year-end full count, both stores: snapshot → scan (floor closed, back-room moratorium) → Kyle applies the difference. Supersedes INVENTORY_COUNT_DESIGN.md for counting.
- [COUNT_OFF_SPEC.md](project-fawkes/docs/COUNT_OFF_SPEC.md) — **NEW 2026-10-30, spec only.** Staff report "Shopify says M, there are N" from the Restock row or Scan card; Kyle alone applies it to Shopify (`count-act`, `inventorySetQuantities` + `compareQuantity`). Born from the Campari count hunt (log 2026-10-30).
- [CORPUS_STANDARDS.md](CORPUS_STANDARDS.md) — the editorial quality bar ("unimpeachable"), six-tier sourcing framework, author model ("Kyle authors, AI accelerates"), and the **open** fingerprint-audit plan for the wine-knowledge corpus (the moat — fingerprints/producer notes/appellations powering Tell + Fawkes + Lewis & Clark). Harvested 2026-05-27 from the archived strategic handoff. Engine mechanics are separate (`Project_BlindTasting.md`).
- [log.md](log.md) — append-only chronological event stream. New entries at top. **Read with a line limit** — whole-file reads blow the 256KB tool ceiling.
- [log-archive.md](log-archive.md) — rotated older log.md months (first rotation 2026-06-11: May 2026 and earlier). Same format.
- [index.md](index.md) — this file.
- [PROJECT_LEWIS_AND_CLARK.md](PROJECT_LEWIS_AND_CLARK.md) — long-arc app vision (Tell + TAT + Shop unified). Codename for the wine-intelligence layer. Read when working across surfaces or designing the corpus; skip for purely operational sessions.
- [PROJECT_WALLA_WALLA_WINE.md](PROJECT_WALLA_WALLA_WINE.md) — **NEW 2026-07-11. Concept, committed.** The `wallawalla.wine` vineyard atlas — the concrete *local highway* of Lewis & Clark. Map-first, real AVA polygons → vineyard parcels → auto-derived designate-wine rollup → honest "Somm's Take"; self-maintained via FOMO + open-add + a **refute button** (Community-Notes model; the refute *is* the ego-check). Competitive read: Patti's site / sonoma (squatted) / napavalley (winery-level booster) / Vinous (static) all thin exactly where the corpus + geometry + honest voice win. Same corpus as the Tell, different verb (browse vs deduce) — a ground-truth funnel for it. Geometry already exists in thief-wine-atlas. Mockup: artifact `f205d368`. **Data-parking pipeline live 2026-08-18** → [wallawalla-wine/README.md](wallawalla-wine/README.md) (sources, run/resume, layout; derived report at `wallawalla-wine/data/derived/REPORT.md`). Site code: none yet. **2026-09-04:** §Lane map added (what talks to what); `pipeline/08-coverage-map.mjs` → `data/derived/winery-coverage.csv`; ultra prompts [wallawalla-wine/ULTRA_WINE_CATALOG_PROMPT.md](wallawalla-wine/ULTRA_WINE_CATALOG_PROMPT.md) (atlas wines layer; adapters first) and `thief-wd-rescue/ULTRA_CORPUS_LINKAGE_PROMPT.md` (corpus linkage, nested repo).
- [wa-vineyards/REPORT.md](wa-vineyards/REPORT.md) — **NEW 2026-09-09.** Statewide Washington vineyard census: **579 vineyards, 264 with grower-verified plantings, 150 with coordinates, 23 AVAs profiled**, plus the full Rocks District association roster (32 vineyards). Built because WW wineries source fruit statewide, so the valley roster can't resolve their fruit. Four layers kept deliberately apart (label-inferred / grower-verified / AVA context / OSM geometry) — master file `wa-vineyards/data/derived/WA-VINEYARD-CENSUS.csv`, artifact `6193135c`. Files only; **nothing promoted to canon** — the `vineyards` entity is still a sitting with Kyle. Verifier: `pipeline/05-verify-claims.mjs` (re-reads every cited page; 1,154 of 1,259 claims verified, 93.7% of readable ones held up). Entity migration prepared + rehearsed, awaiting Kyle: `db-migrations/2026-09-09-0001-vineyards-entity/`.
- [zanzibar/bench/VERDICT_PHASE2.md](zanzibar/bench/VERDICT_PHASE2.md) — **NEW 2026-08-25. The Mac Studio memory verdict: 64GB, not 128.** Phase-2 bench (August re-roster + NEW vision lane on 17 real scans): memory scaling inverted — Qwen3.8-27B (fits the owned mini) beat both 120B-class models on text extraction AND was the best vision extractor on the board; 128GB = +$1,200 for measured-worse capability, and Apple's $4,499 config breaks the ~$3k walk-away gate regardless. Carries the graded tables, the buy-gate mapping, and the deferred lanes (coding, Claude division, grind-share).
- [zanzibar/bench/VERDICT_PHASE3.md](zanzibar/bench/VERDICT_PHASE3.md) — **NEW 2026-09-26. First local inference on the Studio + NVIDIA free tier.** Gemma 4 31B leads (80% hosted / 77.5% local extraction); MLX runtime + run commands; the prompt-cache GPU out-of-memory landmine and its flag fix.
- [zanzibar/bench2/WHY_256GB.md](zanzibar/bench2/WHY_256GB.md) — **NEW 2026-11-06.** Shareable summary of the Nov 1-5 bench2 + quant bench that made the 256GB M5 Ultra Studio the box (128GB = worse tier, not cheaper). Written to paste into claude.ai; supersedes VERDICT_PHASE2's 64GB call for the big-model question.
- [zanzibar/OVERNIGHT_RUNS.md](zanzibar/OVERNIGHT_RUNS.md) — **NEW 2026-10-27.** Plan for three Studio overnight brute-force runs (wine fingerprints, Hanafuda self-play, grounded product copy): one per night, plain scripts, no prod writes, named verifiers. Nothing built yet.
- [PROJECT_ZANZIBAR.md](PROJECT_ZANZIBAR.md) — **Benchmark phase 2026-07-22; Phase 1 RAN 07-22/23, Phase 2 RAN 08-25 (see VERDICT_PHASE2.md above).** The shop brain + local-LLM question, decoupled: brain = owned M4 Pro mini; AI tenant gated on a grounded real-work benchmark (7-model roster, faithfulness-to-source metric, ~$8–15, Kyle gates spend). Contains the dispatch-ready Phase 1 brief, the July-2026 hardware landscape (RTX Spark fall box, M5 Ultra rumors, DRAM-crisis GPU carnage — sourced), and the non-myopic buy gates. Fall decision window by design.
- [docs/ideas/ROOMS_NOT_FEEDS_BOOK.md](docs/ideas/ROOMS_NOT_FEEDS_BOOK.md) — **SEED 2026-10-22, not a commitment.** Book/essay idea: "a room is a thing you can be a regular in"; rooms vs feeds; scenes named; essay first if it still tugs sober.
- [docs/ideas/PROJECT_LASER_BOXES.md](docs/ideas/PROJECT_LASER_BOXES.md) — **NEW 2026-07-18. Concept, pre-sell gate open.** Engraved wooden wine boxes fulfilled locally (B2B small runs for wineries = the engine; retail customization = marketing layer). Machine research done (CO2, Thunder Nova 35 top pick ~$9-13k; Z-clearance for assembled boxes = the killer spec), install + economics + pitch draft inside. Nothing bought — gated on 2-3 winery pricing conversations. Thief-adjacent; brand/entity fork deliberately open.
- [parlor/CHARTER.md](parlor/CHARTER.md) — **NEW 2026-10-23. Committed, Phase 0.** The Parlor (working name): one lobby, many games (hearts/spades/cribbage/dominoes/hanafuda/mahjong/word game...), rooms not feeds, bots at five levels in every empty seat, no chat by default, no money on the table, iOS+Android+web from one codebase. Kyle's gift to the internet; hanafuda is the first door. Laws, game table, 5-line architecture, cost ladder, phases, six decisions for Kyle. Session rules + pending: `docs/rules/parlor.md`.
- [hanafuda/PROJECT_HANAFUDA.md](hanafuda/PROJECT_HANAFUDA.md) — **NEW 2026-10-21, UNPARKED same day.** Hanafuda app: Koi-Koi + Go-Stop + Hawaiian sakura as equal co-flagships on one engine, basic bots + history layer = step 1; no release until complete; physical deck = Phase 2. Labor of love, not a product bet. Rulings on top of the file.
- [hanafuda/SPEC.md](hanafuda/SPEC.md) — **NEW 2026-10-23.** The three rule sets side by side, derived from the engine so it cannot drift; §6 = the gavel sheet of every ruling Kyle still owes (22, pre-filled yes/no). Verifier `hanafuda/scripts/verify-spec.mjs`. Siblings: `hanafuda/RULES_NOTES.md` (sources + flags), `hanafuda/engine/BOTS.md` (two bots, bench numbers, open gate), `hanafuda/content/encyclopedia/` (12 cited month pages + `values.json`).
- [PROJECT_MOTHER_GOOSE.md](PROJECT_MOTHER_GOOSE.md) — **NEW 2026-07-27. Concept draft, NOT ratified.** Golden Goose productized for other Shopify wine shops ("receiving for Shopify retailers"). Born from the Santé recon: Santé's payments-based model can't serve Shopify-native shops = counter-positioned wedge. Validation ladder (gut-check → paid concierge pilot → only then code), kill criteria, metered-COGS caveat inside. Next move is Kyle's: name 3-5 shop owners for rung 2.
- [SENSE_AND_ANCHOR.md](SENSE_AND_ANCHOR.md) — **NEW 2026-08-12. Relationship dossier, not a project.** Kyle's buddy/best-customer's wine-journal app (blacktonguesociety.com): ALLY stance, advisor-offer + stock options on the table, the "he wants my data" four-pile framework (customer=never · transactional=no · corpus=commercial-license-only, guards the L&C v2 asset · catalog-feed=plausible yes), revenue-model watch ("who pays, in one sentence"). Update after every founder conversation so Kyle never retells it.
- [PROJECT_SHOPIFY_B2B.md](PROJECT_SHOPIFY_B2B.md) — native Shopify B2B replacing B2B Wholesale Hub (Industry/Employee/Family tiers). Phases 1-3 LIVE: catalogs seeded (13,242 fixed prices), catalog→mirror sync on cron, GG authoring wine+beer tiers at ingest. **Open: spirits tiers** — must target the bundle *child* (kill-the-bundle stood down 2026-06-12); plus customer linking for online B2B go-live. ⚠ Contains a self-flagged stale claim ("zero enrolled companies") — re-verify before trusting. Read before any B2B/tier-pricing work.
- [PROJECT_HORIZON_2026.md](PROJECT_HORIZON_2026.md) — **H2-2026 stage-setting** (Fable, 2026-07-07): constraint set, Q3-turn-on/Q4-compound arc, 15 executable blueprints (B1-B15) + 11 weird-wall ideas (W1-W11) with kill criteria + anti-goals. The answer to "what should we focus on." Reviewed/reshaped at each quarterly `/retro`; log.md outranks it.
- [shopkeeper.md](docs/rules/shopkeeper.md) — **NEW 2026-10-12.** The bottle-shop Shopify app (OND Lane 6, Kyle-ruled 10-10): flat fee, first feature = multi-tenant Shopify→Xero daily sync; repo `~/code/shopkeeper/`; the 7-day Thief fixture diff is the contract; hard stops until lifted.
- [docs/rules/wallawalla-wine.md](docs/rules/wallawalla-wine.md) — **NEW 2026-10-03.** wallawalla.wine working order: measured state, readiness verdict (Phase 1 shippable; blockers are Kyle's go-live call, one design sitting, the stars sitting), 9-row order of work, open rulings. Vision stays in `PROJECT_WALLA_WALLA_WINE.md`.
- [PROJECT_WASHINGTON_BOOK.md](PROJECT_WASHINGTON_BOOK.md) — **NEW 2026-07-30. Kyle intention, nothing scheduled.** The Washington wine book, modelled on Bonné's two-volume *New French Wine* and against *Vino Italiano* (single block → aged all at once). Thesis: **the narrative ages like a book, the directory ages like a database** — so the directory half is a *render*, not a manuscript, off a living corpus. Why it doesn't exist: reference literature is a byproduct of national distribution, and WA opted out. Contains the verified asset inventory (849 vendors probed / 109 scraped / 4,380 rows) and the dual-purpose finding — **the corpus work already queued for the menu IS the book's sourcing.** Inherits the wine-attribution discipline at maximum severity: print freezes errors.
- [DREAM_THE_CELLAR.md](DREAM_THE_CELLAR.md) — **NEW 2026-08-01. Dream memo, nothing committed.** Opus + 6 web scouts, on a Kyle "dream something magical / dare me" ask. **The finding that outranks the ideas: 145 customers = 52.5% of revenue, flat ~50% every year 2018-2024 — and 95 of them have no address on file.** Three ideas: **The Middle** (RNDC Ch.11 filed 7/26/26 + Columbia absorbed the OR/WA book + SGWS cutting reps ⇒ a wholesale-service vacuum, walkable via **RCW 66.24.179** reseller endorsement whose 24L cap HB 1636 removed 1/20/26 — ⚠ verify the endorsement + RCW text first) · **The Cellar** (sell *time*: pre-populated cellar ledger + drinking-window clock for the 145; **build it backwards — ledger/clock pilot needs zero capital**, the room only if it validates; ⚠ CellarTracker's CellarChat is partial prior art, the moat is narrower than "nobody has this") · **The Indie Panel** (pooled aggregate-only sell-through across independent shops — the channel Circana/NIQ don't cover; sell the aggregation, never the lake). §7 carries the adversarial teardown honestly, including the item that beats every idea here: **NS has been open 5 months with no grand opening, in the only growing category (RTD +16.4%)**. ⚠ 2025 whale-share (67%) is a WD→Shopify migration artifact — do not cite. **§11 = falsifications actually run, two failed:** the corpus has **no drinking-window data** (`corpus_vintages.aging` is cooperage; `vintage_charts` = 226 rows/19 regions; sizing = 551 producer profiles, not 3,465 bottle calls), and **`cellar_entries` already exists** — schema + RLS + SwiftUI, parked out of Fawkes 2026-04-24 for TAT, *parked not abandoned, don't delete*. Contact coverage on the 145 corrected: **89% email, 31% address** (gap is physical mail, not reachability).
- [WEIRD_WALL_IDEAS.md](WEIRD_WALL_IDEAS.md) — **NEW 2026-07-28. Idea bank, nothing committed.** 12 off-the-wall shop projects from a Kyle ask, each with what-it-takes + a kill criterion, tiered A/B/C. Kyle: "all great with a few tweaks" — but he hasn't named the tweaks, so **nothing here is buildable off this file alone**. Someday card on the board; pluck one at a time. Dedupe notes against Horizon W1-W11 inside (B2≈W5, C1≈W3). The long-form companion to `PROJECT_HORIZON_2026.md` §Tier 4.
- [PROJECT_STAFF_PICKS.md](PROJECT_STAFF_PICKS.md) — **NEW 2026-07-28. Gameplan, nothing built.** Staff picks on thiefshop.com — and specifically the buy-in problem Kyle named. Core move: **nobody writes anything**; Kyle interviews each staffer 10 min on the floor, the transcript becomes a draft, they only edit. 3 picks each (under-$25 / splurge / wildcard), quarterly, auto-dark at zero inventory. Surfacing verified against the live Vantage theme: `product-listing.liquid` already reads metafields (the reviews-rating row) — badge + subheader anchor there. ⚠ Smart collections can't filter metafields → tag AND metafield, written together. Ends in 4 Kyle rulings.
  - [marketing/staff-picks/2026-10-17-drafts.md](marketing/staff-picks/2026-10-17-drafts.md) — round-1 pick drafts from the 10-17 floor interviews (Matt, Flint, Henry, Dan); awaiting Kyle red pen + staff approval.
- [wine-dream-round2-for-codex.md](wine-dream-round2-for-codex.md) — cross-vendor strategy loop (2026-07-10): Fable's grading of Codex Sol's blind wine-data dream, the round-2 ask, and the Round-3 addendum (4 questions: living-inventory v0 falsification, marketing steelman, one intervention type, Tell v1 consent tiers). Codex reads this file directly; he must NOT read the wiki — the outside view is the asset. Round 2 moved doctrine: sensory dataset = highest-ceiling asset (concession), rights three-field model, capped-trial rule, B15.
- [WINE_DREAM_SYNTHESIS.md](WINE_DREAM_SYNTHESIS.md) — the **internal** companion to the Codex memo (2026-07-11): full doctrine ledger with reasoning decompressed (sensory-asset ranking, three-field rights, capital-as-test, behavioral laboratory, occasion signal, B15 design laws), the data-dig grounding in nine beats, Fable's sealed round-3 answers (collide with Codex's, don't echo), and the collision protocol. Never show to Codex.
- [inventory-intel/B15_BLUEPRINT.md](inventory-intel/B15_BLUEPRINT.md) — buildable spec for Horizon B15 (living inventory + buyer's prep): substrate audit (reorder engine live but loop unowned — 470 queue rows all `new`), 12 verified data findings, v0 falsification queue (`inventory-intel/recon-2026-07-11/v0-queue.html`, 27 graded positions), 4 build phases, design laws. Umbrellas `docs/rules/reorder-system.md`.
- [TheTell/CAPTURE_SCHEMA_AUDIT_2026-07-10.md](TheTell/CAPTURE_SCHEMA_AUDIT_2026-07-10.md) — Tell capture-layer audit vs the Codex Part-5 "capture now" list. Central finding: the app persists NOTHING (no answers/sessions/identity/reveal-scoring) — greenfield, so this doubles as the capture-schema requirements spec. v1 ratify list of 6 + two Phase-C spec gates (unassisted lane + AI-output logging). Read before any Tell answer-tracking or consent work.
- [TheTell/PR7_CAPTURE_SCHEMA_BRIEF.md](TheTell/PR7_CAPTURE_SCHEMA_BRIEF.md) — executor brief for the capture layer (new track, parallel to Phase A): 5 INSERT-only tables (sessions/answers/final_calls/reveals/ai_events) + `truth_class` on nodes, anon-INSERT-no-SELECT security model (no new definer fns), lock-in + truth-step flow, outbox/never-halt writes, build flag gated on consent. Ten design decisions made; ⚠ FORK = truth_class backfill CSV (Kyle ratifies). Gated on the capture-audit ratify card.
- [TheTell/CONSENT_TIERS_FABLE_DRAFT.md](TheTell/CONSENT_TIERS_FABLE_DRAFT.md) — Tell v1 consent design, Fable's half of a two-draft collision (SEALED before Codex round-3 Q4 arrives; neither ships alone). 3 tiers: local-only play / one bundled in-house contribute choice / external lane Settings-only default-off. Flags a real bug: the "your palate never leaves your phone" tagline vs server-side capture. **Superseded by the ratified merge below — historical.**
- [TheTell/CONSENT_TIERS_MERGED_2026-07-15.md](TheTell/CONSENT_TIERS_MERGED_2026-07-15.md) — **the consent design of record (Kyle-RATIFIED 2026-07-15)**: Fable draft × Codex round-3 reconciled. Tier-0 local play · core cloud disclosure · improve-the-product (asked after first tasting) · commercial research (asked after several sessions). Carries the tagline law. Clears PR-7's second gate; supersedes both parent drafts.
- [PROJECT_WINE_CORPUS.md](PROJECT_WINE_CORPUS.md) — wine knowledge corpus (the data layer powering Atlas / Tell / TAT / Fawkes). **✅ BUILT AND LIVE 2026-07-11:** `corpus_*` tables + `build-corpus.mjs` union of all 13 importer silos + distributor book (58k wines, per-field source attribution), `v_surface_wines` materialized front door, tasting skill + menu-tool wired. Status header lists Kyle's open backfill gates; Appendix A (the last-inch-never-walked pattern) stays as the cautionary record. Lives inside the `CORPUS_STANDARDS.md` contract.
- [WINE_CORPUS_SURVEY.md](WINE_CORPUS_SURVEY.md) — **NEW 2026-07-11.** The measurement layer for the corpus: every column in all 13 importer silos, what it means, what grain it sits at, plus live link rates (**22,521 wine rows, 396 linked — 1.8%**) and the country-backfill scoping. Read for *what we own*; read `PROJECT_WINE_CORPUS.md` for *what to do about it*.
- [CORPUS_SCHEMA_DESIGN.md](CORPUS_SCHEMA_DESIGN.md) — the locked corpus structural model (2026-06-11): three layers (raw staging / curated reference / derived canonical), "polaroid" rebuild-verify-swap build contract, table sketches (corpus_wines / corpus_vintages / corpus_prose / corpus_scores), conflict-resolution precedence. LOCKED 2026-06-11 (red-pen items resolved inline) — read before any corpus migration or landing work.
- [CHAIN_OF_CUSTODY.md](CHAIN_OF_CUSTODY.md) — **NEW 2026-06-14.** The identity-spine model: vendor→producer→importer-of-record→distributor, with the rule *who you buy through ≠ whose content is good* (chain of custody vs prose source — the scrape conflated them in `wine_producers.importer_id`). Two-sided + graduated tiers (distributor catalog `distributor_products` is PRIMARY since the importer spine is sparse). Resolver: `thief-wd-rescue/scripts/resolve-chain-of-custody.mjs` (read-only → review CSV). Read before vendor de-pollution / producer↔importer↔distributor work.
- [DISTRIBUTOR_LIVING_CATALOG.md](DISTRIBUTOR_LIVING_CATALOG.md) — **NEW 2026-06-15.** The temporal / change-detection layer ("jazz singer, not flip-book") over distributor scrapes: surfacing price drops / restocks / discontinuations / new arrivals *between* scrapes. 3-tier identity (episode/format/lineage), SGWS status vocabulary (`DP`=Discontinued Item Number) + productDetails discontinuation trio, **change-detection keying rules** (pair on `ec_prd_itemnumber` NOT `ec_skus[0]`; per-bottle price; UPC for shelf-join only), buy-list spine (`distributor_invoices`/`_lines`), warehouse=dead-end, the Chartreuse 750→700 live test case. Layer 2 lifecycle resolver = next build. Read before any distributor change-detection / ongoing-scrape / reorder-signal work.
- [DATA_CARD_SPEC.md](DATA_CARD_SPEC.md) — **NEW 2026-06-12.** What "complete" looks like for every Thief product (wine/beer/spirits): the field→tag→`custom.*`/`shopify.*`→source-precedence card per type, + the assemble→diff→push engine (self-heal + self-create). The blueprint the enrichment pipeline builds against; supersedes the net-new-only scope of `golden-goose/DISTRIBUTOR_FIRST_ENRICHMENT_DESIGN.md`. Holds the data-warehouse framing (corpus = warehouse, Shopify = client #1) + Kyle's locked decisions + the gaps build list. Phase-1 (importer + distributor geo/varietal/notes into net-new) SHIPPED 2026-06-13.
- [ANALYTICS_SPINE.md](ANALYTICS_SPINE.md) — **NEW 2026-07-16.** One sales truth across both eras: `wd_sales_master` (frozen) + `shopify_sales_master` (to build — sync currently drops customer/staff/guest data) + `v_sales_unified` view. The two-lane seam model (Shopify rang Thursday tastings for YEARS pre-cutover; WD legacy card-on-file tail post-cutover → no dedup needed, but **per-system YoY is dishonest — union only**). Metric views (YoY, capture rate w/ WD-2025 baseline 55%/$ · 35%/orders), surfaces in order (HTML v0.1 → Power BI [Kyle has M365, $0] → Fawkes staff scoreboard), incentive-bonus sequencing law (baseline → scoreboard → money; payroll regular-rate flag). Provenance-marked [K]/[P]. Read before any analytics/dashboard/KPI/incentive work. (~$1K/yr) with native Shopify B2B (online, 3 catalogs via B2B markets — Grow constraint) + an own POS UI extension (tag→`setLineItemDiscount`). Tier model (retail/industry/employee/family, validated on live data), 3-bucket migration (seed/fix/engine-fill 3,280 gaps), and the GG pricing engine that fixes margin erosion. SPEC phase, committed 2026-06-02.
- [robin-hood/](robin-hood/) — standing OSS contribution practice (the "Robin Hood" project). Three files inside: `README.md` (manifesto: triggers, constraints, workflow, loop-closing rules), `LEDGER.md` (open PRs + state, with next-check dates), `CANDIDATES.md` (queued work not yet started). First contributions filed 2026-05-28 at rjhalvorson/skylight-mcp (issue #6, draft PRs #7 + #8). Read before any "ship a fix / contribute / commons work" prompt.
- [command-deck/ROADMAP.md](command-deck/ROADMAP.md) — **NEW 2026-06-02.** Native macOS dashboard for Claude Code usage/leakage (own repo `~/code/command-deck/`; Swift package `CommandDeckCore` engine + `CommandDeckUI` SwiftUI face, plus a CLI harness). Reads local `~/.claude/projects/*.jsonl`, **zero Claude calls, $0 to run**. Engine verified ±2% vs ccusage; ships per-project breakdown + MCP dead-weight detector + live auto-refresh. Vision/roadmap + the billing model (cockpit embeds *interactive* `claude` = subscription bucket, not metered) live in ROADMAP.md.

- [docs/archive/2026-09/WWDC26-notes.md](docs/archive/2026-09/WWDC26-notes.md) — **NEW 2026-06-09.** Remote-watch scouting notes mapped to Thief. Headline: Foundation Models can front Claude via the `LanguageModel` protocol + an on-device `fm` CLI (macOS 27). Two corrections banked (menu-tool `-p` was never metered; reaching FM doesn't need a native rewrite) + Kyle's truly-free vs prepaid-bucket-free distinction. Ranked action items; top = fall-2026 `fm`-in-`voice.ts` spike. Decide if in-person WWDC/Claude-conf earns airfare next year from what this yields.

- [docs/ideas/HOTEL_COOP_FEASIBILITY.md](docs/ideas/HOTEL_COOP_FEASIBILITY.md) — **NEW 2026-07-18. Feasibility-stage / dream (not committed).** Winery co-op-funded hotel ("The Twelve Visits" codename) to reverse WW tourism *frequency* collapse (12 visits/yr → 2): co-op ownership internalizes the lost-visit externality, two-tier rates (capped member/club tier + open market tier), member room-night routing commitments = bankable occupancy floor, dividends subordinate to debt. Napkin DSCR table (30-key conversion, ~$5M), capital classes incl. a customer class, order-of-operations gates (POS frequency proof → 10 winery coffees → Alliance/Port/LTAC-funded feasibility study). Kyle = instigator, not solo builder. Kill criteria inside.

- [docs/ideas/FIRE_COMPLIANCE_FEASIBILITY.md](docs/ideas/FIRE_COMPLIANCE_FEASIBILITY.md) — **NEW 2026-06-15. Feasibility-stage (not a committed project).** A new local business line under eval: commercial fire-extinguisher + backflow compliance service in Walla Walla (kitchen hood as orchestrated add-on). The opening = PE roll-ups (Summit/PSI, ex-Oxarc, Norco) abandoned the small-account long tail; demand is statutory (WW commercial backflow due July 15), entry is ungated in WA (FE $320 / BAT ~$331, no state license), the moat is local presence. Holds verified cert/AHJ/submission-rail facts (WW is paper/email, no portal), the gated-vs-ungated insight, a 3–6wk ops-software read (build-for-self not SaaS-competitor), and a this-week validation-call kit. Posture: self-fund the reference op (Kyle or Beverly certifies), operator-financing deferred past deleverage. Gate: 3 calls + chef's reply → go/park.

## Fable kit (`fable-kit/`) — the portable any-model operating kit

_Built 2026-07-07 (Fable's last day). The OUTSIDE-the-repo leg of the exit package: paste-able context, weaker-model harnesses, and model-routing logic for any AI surface without wiki access. Inside Claude Code the wiki stays authoritative; every kit file carries a decay clause (log.md outranks) and the refresh ritual is PROMPT_LIBRARY #20 (quarterly at `/retro`)._

- [fable-kit/README.md](fable-kit/README.md) — how the kit fits together: the fleet table (Claude Code / Codex / claude.ai / NotebookLM / free tiers) + the 5-pattern multi-model workflow + honest what-doesn't-transfer limits. **Read first.**
- [fable-kit/COMPACT_CONTEXT.md](fable-kit/COMPACT_CONTEXT.md) — the <1,500-word passport (1,258 wc-verified): paste first into ANY new AI chat, any provider.
- [fable-kit/WHATS_NEW.md](fable-kit/WHATS_NEW.md) — the catch-up digest for claude.ai and any outside AI: shipped-to-real-people, data, security, OS changes, and the finish-line table (built, not released). Written 2026-10-07 from log + log-archive (08-07 → 10-07); regenerate when stale.
- [fable-kit/MASTER_CONTEXT.md](fable-kit/MASTER_CONTEXT.md) — the full portable context (~2.5k words) for long/serious sessions off-repo.
- [fable-kit/SPECIALIZED_PROMPTS.md](fable-kit/SPECIALIZED_PROMPTS.md) — the Universal Header (echo-back, UNSURE-don't-guess, scope-lock, self-check) + 10 role harnesses (Extractor, SQL Drafter, Write-Script Surgeon, Code Surgeon, Diff Reviewer, Debugger, Copywriter, Explainer, Log Shaper, Verifier).
- [fable-kit/QUALITY_CHECKLIST.md](fable-kit/QUALITY_CHECKLIST.md) — judging any AI output: 30-second triage, full gate for money/prod/external, slop tells, auto-fail conditions.
- [fable-kit/DECOMPOSITION_MAP.md](fable-kit/DECOMPOSITION_MAP.md) — which model gets which work: 60-second routing triage, Kyle-only/frontier/mid/cheap tiers, never-delegate list, two-strike escalation ladder, worked examples from the H2 blueprints.
- [fable-kit/PROMPT_LIBRARY.md](fable-kit/PROMPT_LIBRARY.md) — 20 ready prompts mapped to July–October 2026 (importer scrapes, enrichment QA, invoices/pricing, marketing content, audits, the kit-refresh ritual).

## Project rule files (`docs/rules/`)

_Project-specific context. Read on demand. Updated at session end._

**Path-scoped rules (`.claude/rules/`, NEW 2026-10-31)** — short pointer files the harness loads automatically the first time a matching file is opened (nothing at session start). They point INTO `docs/rules/`; they are not copies. ⚠ A `paths:` line on a `docs/rules/*.md` file does nothing; only `.claude/rules/` is read.
- [.claude/rules/golden-goose.md](.claude/rules/golden-goose.md) — `golden-goose/**`: read ARCHITECTURE first, three price-math copies, bundle create-only, test gates.
- [.claude/rules/shopify-xero-sync.md](.claude/rules/shopify-xero-sync.md) — `shopify-xero-sync/**`: push = prod, branch→diff→verify, Codex review, 60/min Xero quota.

- [db-atlas.md](docs/rules/db-atlas.md) — **NEW 2026-08-21, DRAFT (awaiting Kyle's veto).** The Kyle-altitude map of Supabase: all 321 relations grouped into **6 neighborhoods** (shop floor · the WD vault · receiving/invoices · sourcing catalogs · the wine library · people+apps), each in 4 lines — what it holds / who writes it / **precious vs rebuildable** / blast radius if wrong. Three-minute read, no column dumps; the orientation layer *above* `CLAUDE_SUPABASE.md` (meaning) and `db-docs/` (structure), both of which outrank it on any conflict. Read it to answer "how bad is it if this breaks" before touching an unfamiliar table.
- [pricing-and-product-rules.md](docs/rules/pricing-and-product-rules.md) — **authoritative home** for pricing math, product-type taxonomy + `\b` regex, vermouth=wine / Pour=Wine, match priority, COGs split, invoice convention. CHEATSHEET + B2B point here.
- [wd-sales-master.md](docs/rules/wd-sales-master.md) — **NEW 2026-07-15.** Build spec for `wd_sales_master` (PROD, line-grained WD source-of-truth: spine + card/tip welded, 311,747 lines / $8.62M / 2018–2025, all sales carded+non-carded). Source registry (sales-detail-x spine / tender card+tip / quickbooks optional / daily-activity recon oracle), layered column map, 4 sharp edges (line-id, aliasing, money semantics, 2025 card blackout), FTI attribution + claim model, open `DECISION:` markers. In-flight: specs layer via fresh admin export (controlled-vocab → canon lookup). Feeds [[loyalty-membership]] Brix seed.
- [BRIEF_RELAY_LINK_CARD.md](ThickAsThieves/BRIEF_RELAY_LINK_CARD.md) — **NEW 2026-09-11.** Opus brief: unmatched Sign-in-with-Apple relay accounts get "Link your shop account" on the Home card instead of a QR; verified-email link (code via Resend) attaches the app to the customer's existing identity row and retires the nightly-minted twin. Kyle's ruling: the register pick is skipped by staff; put the burden on the customer.
- [BRIEF_PASSPORT_PHASE2.md](ThickAsThieves/BRIEF_PASSPORT_PHASE2.md) — **NEW 2026-09-11.** Opus brief, the Passport in the app: Home card → full-screen splash (prototype beat sheet = spec) → SwiftUI vector map (generated world outlines, pins) → place sheet → Shop pre-filtered via `p_place`. Contracts = the live Phase 1 RPCs. No DB change, no journeys.
- [BRIEF_PASSPORT_PHASE1.md](ThickAsThieves/BRIEF_PASSPORT_PHASE1.md) — **NEW 2026-09-10.** Executable Opus brief, Passport data spine: `passport_places` (path slugs) + aliases + Barolo MGA table, the one resolver, `v_passport_placements`, `get_my_passport()` / `get_my_passport_place()`, `p_place` on `get_shop_products` (overload cleanup), the Shopify-only-customer Cellar fix, coverage report with three controls. No app code.
- [passport/WISHLIST.md](ThickAsThieves/passport/WISHLIST.md) — **NEW 2026-09-16, rewritten same night.** Kyle's v2 map wishlist, now mostly answered: grand crus need **no work** (they are in EU-PDO and in our canon — unseeded only because the seed is scoped to places we have SOLD from, so they seed themselves the day he sells one); Barbaresco's 66 MGAs + 4 communes are **landed and cited**, with seeding gated on a prod migration and a fifth tree level; Australia **done**; the Barolo pin question **closed** (the counter-scale has a ceiling). The one live open item is §6 — the map drifts ~59pt at extreme zoom, and the fix is Kyle's call.
- [passport/IGP_PLAN.md](ThickAsThieves/passport/IGP_PLAN.md) — **NEW 2026-09-23.** France's 76 wine IGPs: INAO commune-list source (Licence Ouverte), parents by commune majority, measured demand, 5 pre-filled rulings, build order (the Champagne recipe).
- [passport/DONE_LIST.md](ThickAsThieves/passport/DONE_LIST.md) — **NEW 2026-09-23. THE ship gate.** Kyle's ruled definition of done: 7 items (phone walk, honest empty map, rough edges, morning-after push, two journeys, IGPs, same-day stamps) + release steps. New asks go under After launch unless Kyle moves them up.
- [passport/reviews/2026-10-05-codex-architecture-round-1.md](ThickAsThieves/passport/reviews/2026-10-05-codex-architecture-round-1.md) — Codex blind ARCHITECTURE review of the whole Passport before 1.11 (SHIP WITH FIXES; the five fixes + the overrule are in log 2026-10-05). First whole-feature read; the piecemeal-build answer.
- [passport/COMPLETION_LEDGER.md](ThickAsThieves/passport/COMPLETION_LEDGER.md) — ruling 19: the 100 unshaped seeded places, six buckets, order of work (2026-09-16)
- [passport/MAPLIBRE_MIGRATION.md](ThickAsThieves/passport/MAPLIBRE_MIGRATION.md) — **NEW 2026-09-24. The map-engine migration plan of record** (Kyle ruled MapLibre Native): what the SDK deletes vs what stays ours, decisions D1–D8 pre-filled, five sequential slices with verifiers, the spike's silent-failure traps. Read before touching the Passport map.
- [passport/MAP_PLAYBOOK.md](ThickAsThieves/passport/MAP_PLAYBOOK.md) — **NEW 2026-09-15.** The map's playbook: pipeline diagram, copy-paste recipes (add a country / river / landmark town / 1:10m outline), a table of every generator control WITH the real failure that created it, and the full landmine list — Natural Earth's own misspellings (`Sane`, `Rhin`, `Mio`), `topojson.merge()` not being a polygon union, rings vs lines in Douglas-Peucker, lattice-vs-float chaining, `BUILD SUCCEEDED` with no binary, text that cannot be counter-scaled through a zoom, and (2026-09-16) **the counter-scale's own ceiling** — it stops working on PINS and strokes somewhere above a region page's zoom — plus a country-scale layer gated in its fill but not its stroke. **Read before touching map geometry.**
- [passport/README.md](ThickAsThieves/passport/README.md) — **NEW 2026-09-10, extended 2026-09-11.** The Passport place spine: the `data/*.csv` artifacts + every script that rebuilds them, the splash-animation prototype, and (phase 2) the two Swift generators — `gen-swift-world.mjs` (Natural Earth outlines + the projection ported to Swift) and `gen-swift-seed.mjs` (the 571 seeded places, so the map can grey out where you haven't been). Each carries its control.
- [PASSPORT_JOURNEYS.md](ThickAsThieves/PASSPORT_JOURNEYS.md) — **NEW 2026-11-14.** Journeys round two (design, nothing built): Kyle's rulings (right before stocked, trail ratings, producer + versus journeys, out of the Swift files, versus vote), six kinds, the one counting rule, server tables + functions, the facts pipeline (`passport/journeys/` dossiers + verifier), the slate by trail, order of work.
- [PASSPORT_ENGAGEMENT.md](ThickAsThieves/PASSPORT_ENGAGEMENT.md) — **NEW 2026-09-16.** The engagement layer on the passport (addendum to PASSPORT_DESIGN): measured usage table, M1 stamp moment (in flight) · M2 stamp push · M4 standing · M6 share card · M7 staff gaps · M9 Wrapped (scope mid-Oct); order of work + four rulings pre-filled for Kyle. Materialised placements is the prerequisite for every server-side mechanic.
- [PASSPORT_DESIGN.md](ThickAsThieves/PASSPORT_DESIGN.md) — **NEW 2026-09-10.** Design sitting (not a brief): scratch-off Passport map + guided Journeys painted from `customer_wine_history`, inside the Cellar tab. Seven rulings pre-filled for Kyle; live coverage numbers; the 5-line data map; the Rhône journey worked as the rule-language proof; three sequential phases. Becomes the brief on Kyle's Y.
- [BRIEF_PRODUCT_DETAIL.md](ThickAsThieves/BRIEF_PRODUCT_DETAIL.md) — **NEW 2026-09-06.** Executable Opus brief, scan-to-bottle: `get_shop_product_detail(handle)` RPC (list fields + bio/notes/importer/scores/corpus soil-farming; never `staff_notes`) → iOS/Android product-detail sheet built for the half-populated warehouse (sections render only when present; coverage table inside) → `go.thiefshop.com` AASA/assetlinks + associated domains so shelf QRs open the bottle in the app. Kyle: "the shop doesn't have drill down… worthwhile" + "the data is a lot better."
- [REVIEW_2026-09-05.md](ThickAsThieves/REVIEW_2026-09-05.md) — **NEW 2026-09-05.** Fable night review of TaT: what shipped (upcoming events → blog post link: migration `db-migrations/2026-09-05-0007-event-blog-link`, iOS + Android rows, publisher stamps `blog_url`), the open-work inventory that already existed (Your Somm server-live/UI-unbuilt at the top), six ranked next builds with my call pre-filled, and the icon direction (purpose-drawn sunburst mark per Kyle's 07-25 ruling; tab glyphs). Two Your-Move cards point here.
- [SOMM_TAB_BRIEF.md](ThickAsThieves/SOMM_TAB_BRIEF.md) — **NEW 2026-07-17.** Executable Opus brief for the TAT "Your Somm" tab: renders live `get_my_somm_dossier()` (server side DONE), empty-state-first design law (the App Store reviewer sees it), house-voice pass Kyle-gated, hosts the allocation Drawing entry if ratified. Post-v1.0-approval update.
- [DROPS_RUNBOOK.md](ThickAsThieves/DROPS_RUNBOOK.md) — **NEW 2026-07-24.** The operator's manual for the drops machine: the state machine, the four verbs (`get_drop_candidates` → `create_drop` → `open_drop` quiet / `fire-drop-wave` loud), the guardrails, **what does NOT exist yet** (nothing closes a drop, a claim notifies nobody, no cron touches drops), and copy-paste prompts for a future session. Read this before touching drops.
- [DROPS_NEW_VINTAGE_BRIEF.md](ThickAsThieves/DROPS_NEW_VINTAGE_BRIEF.md) — **NEW 2026-07-24.** New-vintage drop pipeline execution brief (v1 manual-with-preview → automation ladder); Kyle's 3 decisions flagged; drop-#1 candidate = Chiarli Sorbara 2024 for Beverly. Written launch night; sitting carded. **Automation ladder SUPERSEDED 2026-08-05 by the vintage-flow spec below (Kyle ruled: no per-wave gate).**
- [2026-08-04-tat-vintage-flow-v1.md](docs/superpowers/specs/2026-08-04-tat-vintage-flow-v1.md) — **NEW 2026-08-05, SHIPPED in shadow.** Vintage Flow v1: the TAT journey-builder rail — nightly automated personalized "your wine is back / new vintage" drops (arrival-gated ≤30d, exact-title matches only, 60d cooldown, cap-respecting), shadow→live→paused mode switch in `tat_flows`, Monday digest. Carries the claims-to-cash ledger note (stash has no payment teeth — future arc). Operator view: `ThickAsThieves/DROPS_RUNBOOK.md`.
- [2026-08-05-tat-payments-plan.md](docs/superpowers/specs/2026-08-05-tat-payments-plan.md) — **NEW 2026-08-05, RESEARCH+PLAN awaiting Kyle's lane ruling.** The whole TAT payments picture, live-verified: Lane 1 = Checkout Sheet Kit over the proven cart rail (native sheet, Shop Pay/Apple Pay, Functions ⇒ spirits work, 0% Apple per 3.1.3(e)); Lane 2 = native house accounts (store credit prepaid + B2B vaulted-card tabs, all-plans since 2026-04); Lane 3 (Customer Account API in-app saved cards) parked. Rates: app 2.7%+30¢ vs POS 2.5%+10¢. Don't-build list: Stripe sidecar, headless checkout (`cartSubmitForCompletion` docs gone), IAP.
- `marketing/rickhouse/2026-09-16-rickhouse-01-draft.md` — Rickhouse issue #1 worksheet (facts + sources, audience, build record). Issue HTML beside it.
- `marketing/briefing/2026-09-16-briefing-06-draft.md` — Briefing #6 worksheet (Chablis ladder / Occhipinti).
- `marketing/briefing/MONDAY_DRAFT_DESIGN.md` — **NEW 2026-09-16.** The Briefing Monday-draft job, half B (the `claude -p` drafting leg): three calls for Kyle, my call pre-filled (laptop home · worksheet-only + templater · thin issue = Pick + Thursday + holiday bit). Half A = `marketing/briefing/inputs-pack.mjs`.
- [IDEAS.md](marketing/geocities-thief/IDEAS.md) — **NEW 2026-10-02.** The hidden 1998 GeoCities parody at thiefshop.com/pages/1998: what's built (real moderated guestbook, Corky the corkscrew helper) + Kyle's loved-it backlog (dial-up button, screensaver, Konami egg, text adventure, chat room) + Corky-as-app-AI-helper parked.

- [briefing-block.md](marketing/tat-app-launch/briefing-block.md) — **NEW 2026-07-24.** TAT app-launch announcement block for the Briefing. **ABSORBED into Briefing #3 and sent** — kept for the copy history; the Sign-in-with-Apple steer in it is SUPERSEDED (see #3's draft worksheet).
- [1_3_SOCIAL.md](marketing/tat-app-launch/1_3_SOCIAL.md) — **NEW 2026-07-26. POSTED to IG + FB + Story the same day.** The app-launch social copy (IG caption + short alt, FB post with live link, 3-slide shot order, Story plan, prepared replies). Carries the framing lesson worth keeping: it was drafted as a *1.3 feature* post until Kyle caught that the app had only ever been announced in one email — **the app existing was the headline, On the Machine was just the hook**. Also the two screenshot traps: never post `1_3_PREVIEW/` (stale, carried the "Scoth" typo) and **never screenshot your own member card** (the QR encodes your email) — use the `info@thiefshop.com` demo account instead.
- [MARKETING_GAMEPLAN_2026Q4.md](marketing/MARKETING_GAMEPLAN_2026Q4.md) — **RULED 2026-09-12** (stamped copy in `email-sms-klaviyo.md` §13; superseded for strategy by the H1-2027 page below, still law for channel rules). The Q4 ruling sheet built on the channel-role brief: channel mechanics (owned vs rented pipes), the anti-annoyance mechanics + a 10-line stampable rules of engagement, 10 forks with my call on each, the automation matrix (runs itself / Kyle kickoff / should stop needing Claude), acquisition ordered by cost (front doors → the 3,200 → GBP → Google → Meta → billboard=no), 90-day sequence. Kyle rules §4; the stamped version becomes doctrine.
- [MARKETING_GAMEPLAN_2027H1.html](marketing/MARKETING_GAMEPLAN_2027H1.html) — RULED 2026-09-21: three fronts (home turf / the road in / the database), built backwards from Dec 24, the till YoY is the scorecard; $291k Nov+Dec 2025 baseline
- [BRIEF_APP_CHANNEL_ROLE.md](marketing/BRIEF_APP_CHANNEL_ROLE.md) — **NEW 2026-09-08.** Measured ground for the channel-doctrine sitting with Fable: reach by channel (email 5,366 / push 87 / SMS 27 against 7,258 known customers = **1.6% app penetration**), the five notification layers and which have NEVER been used (`news` has 54 opted-in people and zero sends), every push ever sent, and five forks with my call pre-filled. Leads with the provenance finding — **`MARKETING_STRATEGY.md`'s channel doctrine was never ruled on** (one ruling stamp in the file, none on the Channel Map). Kickoff prompt: `.ferry/2026-09-08-marketing-channel-doctrine-kickoff.md`.
- [marketing/calendar/](marketing/calendar/holidays.json) — **NEW 2026-09-02.** The marketing holiday calendar: 178 date rules (wine/beer/spirits/cocktail/food/fun/holiday/season/WW-local) with tier + hook → `resolve.py` → CSVs, `calendar.html`, and the Supabase mirror `marketing_holidays` / `marketing_holiday_dates` (2026-2030). Edit the JSON, rerun, reseed.
- [2026-09-content-sheet.md](marketing/calendar/2026-09-content-sheet.md) — **NEW 2026-09-02.** First calendar-bits content sheet: September slate (Cabernet Day bit, harvest post, Southern Rhône tasting on Grenache Day eve, WW Grenache post, Pirate Day board + Oktoberfest wall, Briefing "coming up" blocks), copy drafted from live stock, 5 gavels pre-filled. The template for every month after.
- [2026-09-captions.md](marketing/instagram/2026-09-captions.md) — **NEW 2026-09-02.** Staged September calendar-bit captions (harvest, Southern Rhône, WW Grenache, Oktoberfest/Pirate, Negroni Week) for the weekly veto; two verify flags inline.
- [2026-07-24-briefing-03-draft.md](marketing/briefing/2026-07-24-briefing-03-draft.md) — **NEW 2026-07-24.** Issue #3 worksheet: the deliberate format break (no Pick, no staff trio, no offer — the shop itself is the story), every fact re-verified live, and the campaign/template ids. Copy #2, not #3, when the normal format returns.
- [2026-08-07-briefing-04-draft.md](marketing/briefing/2026-08-07-briefing-04-draft.md) — **NEW 2026-08-07.** Issue #4 worksheet (Fichet Pick, back to normal format): full arrivals shortlist, Kyle's rulings, campaign ids, and the §12 bounce-pass state. SENT 08-07 to 3,210.
- [2026-07-24-tat-1-2-auth-funnel.md](docs/superpowers/specs/2026-07-24-tat-1-2-auth-funnel.md) — TAT 1.2 spec: the never-wired confirm sheet (root cause), self-healing confirm, push-token cache fix. Shipped same night; kept as the record of the launch-night repair.
- [allocation-program.md](docs/rules/allocation-program.md) — **RATIFIED 2026-07-17 eve (Premier+ · monthly · Blanton's first · doc prices; next gate = LCB email).** How to sell allocated bottles WA-legally: paid raffle = illegal (verified vs RCW/WAC primary text), two legal lanes = The List (deterministic spend-ranked allocation via TAT drops/loyalty — no chance element) + The Drawing (free-entry right-to-buy-at-posted-price, entry decoupled from purchase per WAC 314-52-040(1)(b)). Three-price economics (shelf/list/drawing), 67-bottle shelf audit, LCB confirm-email gate before launch.
- [cron-census.md](docs/rules/cron-census.md) — **NEW 2026-07-16 (harness scoped).** Every recurring job in the stack on one table, classified by trigger (clock/event/kyle) × home (supabase/vercel/local) × risk; the cron-edition silent-write law; the **Harness design** section (4 live pg_cron jobs ground-truthed; run-log contract where http_post jobs log from the edge fn; `v_cron_health` view + Resend email watcher = the whole monitor, no SQL-agent; homes decided incl. the shop Mac mini as a deploy target). Scheduled jobs are on the metered-AI bright line.
- [network-install.md](docs/rules/network-install.md) — **NEW 2026-07-20.** UniFi buildout for both Main St stores (invoice US5829567, $6.8k): gear inventory, locked decisions (RAID1, inter-store 2.5G uplink, termination pattern, Etherlighting color plan, Spirits cutover order), remaining shopping picks, drop-map gate.
- [WebsiteModification/BOT-TRAFFIC-2026-09.md](WebsiteModification/BOT-TRAFFIC-2026-09.md) — **NEW 2026-09-07, OPEN.** The automated-traffic investigation: sessions ~200/day → 11-23k/day since Aug 31, beer crawled 9× harder than anything else, everything ruled out (GSC, Wine-Searcher, our apps, Supabase, an attack), the beacon that will name it, Kyle's email to bots@shopify.com, and why country blocking is neither buildable nor worth buying.
- [WebsiteModification/THEME_MAP.md](WebsiteModification/THEME_MAP.md) — **NEW 2026-09-06.** The theme's what-talks-to-what map: surface → file (PDP gallery, product card, cart line item, quickview), the `.box-ratio` / image-ratio / Bug-#1 rendering contracts, the pull → validate → push loop, and the `probe/` rigs to reuse. Read BEFORE any theme change; `docs/rules/shopify-theme.md` is the changelog, this is the map.
- [WebsiteModification/THEME_UPGRADE_NOTES.md](WebsiteModification/THEME_UPGRADE_NOTES.md) — **NEW 2026-07-27.** Vantage version state (live=12.2.0, vendor=12.3.0 w/ color-system break), the hand-mod inventory that must survive any upgrade (stray-line fix, logo-list styles, custom_css, ⚠ age-gate app embed = publish gate), and the upgrade-day checklist. Read before ANY theme update/publish.
- [security-doctrine.md](docs/rules/security-doctrine.md) — **RATIFIED 2026-07-27 (v1.2).** The ONE house security opinion: threat model (ranked), anti-flip-flop secrets rules + exposure ladder (rotate-now / soon / not-an-emergency), zanzibar host policy (ops keys OK, money/tax NEVER), Tailscale-not-LAN addressing, UniFi segmentation slice, steal-response runbook, verified posture. Cross-vendor reviewed, two rounds (Codex: APPROVE WITH FIXES → five final fixes; all applied in v1.2 — credential manifest §3a, rapportd finding, FileVault-stays-ON, check-only Tailscale updates). Codex would gavel v1.2; awaiting Kyle. Terminals never improvise security stances.
- [mini-host-setup.md](docs/rules/mini-host-setup.md) — **NEW 2026-07-16.** Numbered bench runbook to stand up the shop Mac mini as the hidden always-on local-job host: hidden `thiefops` account, Homebrew/Node/git, Tailscale SSH (the remote-debug door), no-sleep, gh auth, and a heartbeat job that proves launchd→script→log→email end-to-end. Kyle-hands, ~45 min; do it live with a terminal watching.
- [studio-migration.md](docs/rules/studio-migration.md) — **NEW 2026-09-22.** zanzibar mini → M5 Max Studio: the move list (15 jobs, secrets, box-only files), the one-box-at-a-time cutover law, phase order desk→rack, and where the 8 laptop jobs go. Commands live in mini-host-setup.md.
- [desk-mini-migration.md](docs/rules/desk-mini-migration.md) — **NEW 2026-09-22, PROPOSED.** Laptop → desk mini after the Studio cutover: one-home model (mini = home, laptop = Tailscale window via SSH/tmux + Screen Sharing), Migration Assistant move, push-sweep table, what MA doesn't carry, Drive for non-code files.
- [customer-profile-v2.md](docs/rules/customer-profile-v2.md) — **NEW 2026-07-16.** Build spec replacing the frozen-March `customer_profiles` table: identity spine (`customer_identity`) + unified sales view (wd_sales_master ∪ cwh-shopify) + `customer_features` (features-not-verdicts, lifetime/t12m windows, segments as a view) + same-shape compat view so shop-tool/Fawkes needs zero changes + `get_my_somm_dossier()` RPC for the TAT personal-somm tab ($0 deterministic picks; Apple-Intelligence-shaped payload). Audit findings + phases + gates inside. Kyle skim → Opus build.
- [holds-and-stash.md](docs/rules/holds-and-stash.md) — **NEW 2026-08-01. Design of record + rebuild brief** for held orders and TAT's "Your Stash". Kyle's framing governs it: the draft-order mirror is an **onboarding bridge for holds that predate the stash**, deliberately disposable; steady state is **Fawkes → `holds` → Rail**, no draft order ever created. Holds `holds`/`hold_items` (line-level collection; keyed on `shopify_customer_id` NOT tat_account, because most held customers aren't on the app; ⚠ `customer_profiles` is a VIEW, unFK-able), the **§3 paid-for constraints** (shipped-binary decode contract — TAT 1.6 (7) means `get_my_reserves()`'s signature is frozen · `drop_claims` inserts fire **2 REAL pushes per row** and disabling the trigger isn't enough, `sweep_unalerted_drop_claims` re-fires it · empty `notes` = a hardcoded restock lie · pours/services are "owed for", not held · ⚠ **a mirror that re-syncs by REPLACING destroys history** when staff trim the upstream draft · `draftOrders(query:)` silently ignores `name:`/`email:` · ⚠ **two Shopify apps share `SHOPIFY_CLIENT_ID` on disk**), the verified **§4 capability map** (POS retrieves admin drafts natively; API drafts probably invisible — UNTESTED; POS cart write confirmed; Checkout Kit preserves Functions; **Apple takes 0% on physical goods**), and **§6 keep-vs-bin** for the from-scratch rebuild. Kyle's recorded ambition: native in-app checkout, **wine only, pickup only**. **UPDATED 2026-08-25 — the stash now TAKES MONEY:** `get_my_payable_hold()` + `pay-my-hold` + the `orders/create` webhook went live 08-24, and TAT **1.9 (11)** adds the Pay button (built, not submitted). New §: the two money facts that are easy to get wrong (a spirits bundle parent's price **already contains** its hidden WSLT child, so components add $0.00; but WA's extra **20.5% spirits sales tax** puts a spirits-heavy hold ~18% above the shown subtotal — hence "Subtotal", never "Total"), plus the stated ceiling that **a client-side paid-marker is not payment idempotency**.
- [loyalty-membership.md](docs/rules/loyalty-membership.md) — The loyalty/membership decision (2026-06-28): **benefits-only, no cash; Brix = status meter** + industry pricing reframed as per-bottle **trade pricing** (sunsetting into tiers; new staff/POS language; privilege not identity; no stacking) + case discount 10→5. **Tier ladder + benefit grid LOCKED 2026-07-01** — 0/1,000/6,000/20,000 Brix, annual + grace year, per-perk cost structure, TAT ship checklist (January reset job = blocker). Holds the parked paid-inner-circle (StewardShip) option, leaky-cost workstreams (card fees / early-pay), and the email-gate reconciliation flag. Pricing math stays in `pricing-and-product-rules.md`.
- [model-routing.md](docs/rules/model-routing.md) — the terminal/worker model+effort card (3 questions), the Opus 5.5 vs 4.8 bake-off ruling (2026-09-22), and the dispatch ledger (one line per launch; shadow runs make over-reach visible).
- [integrations.md](docs/rules/integrations.md) — MCP servers (+ caveats), cross-project Context7 IDs, tool routing (Supabase-MCP-always, gh, Vercel Hobby cron), dev-env (`CLAUDE_CODE_MAX_OUTPUT_TOKENS`).
- [v2-importer-scrape-playbook.md](docs/rules/v2-importer-scrape-playbook.md) — **NEW 2026-06-15.** Reusable bones for building a wine-LEVEL importer scrape (Skurnik HTML + Dalla Terra API are the two reference impls). Recipe (overlap-check → full-dump-first → raw table → parser → match → reachability/land), the lane rule, the SGWS status decode, the gated landing scripts, the overlap query + per-importer worksheet. Kickoff: *"read this, let's do {Importer}."* Live queue/status: `thief-wd-rescue/IMPORTER_CAMPAIGN.md`.
- [design-doctrine.md](docs/rules/design-doctrine.md) — **NEW 2026-07-11 ("dare to be greater").** World-class-UX methodology for every customer/staff surface: Kyle = taste gate on RENDERS (label-v2 desk-session pattern), never ship the first render, 3 distinct directions, named references before pixels, the three one-time decisions (typeface/color/icon), anti-default rail vs Claude's design attractor, frontend-design skill mandatory on UI builds. Owed: TAT teardown via the Curate-teardown lens; open GitHub design-repo survey.
- [voiceink-dictation.md](docs/rules/voiceink-dictation.md) — **NEW 2026-08-06.** Tuning VoiceInk so dictation stops mangling *Klaviyo / Thick as Thieves / LWIN / Côte-Rôtie*. There is NO pronunciation coaching anywhere — only text levers, and **which levers exist depends on the engine** (Kyle's active Mode switched Whisper→Parakeet 7/23, so the Transcription Prompt is inert and the CTC vocabulary rescorer is live — read the effective model off history, never the Settings pane). Engines graded on Kyle's OWN audio: **Turbo 23.5% vs Parakeet 38.9%** proper-noun error → ruling = back to Large v3 Turbo (full, not q5_0) + the 208-token prompt; Parakeet Unified declined (English-only, same no-prompt family). Order: Word Replacements (always works, was empty) → Turbo + prompt → Vocabulary (test it; needs ASCII aliases for accented terms). 19 measured replacement rules off 468 real dictations. Re-mine with `scripts/voiceink-mishears.py`.
- [build-workflow.md](docs/rules/build-workflow.md) — the 6-step inline-first build default + the price-the-mode template + overnight preflight. (Was `superpowers.md`; plugin uninstalled 2026-08-13.)
- [subagent-prompts.md](docs/rules/subagent-prompts.md) — what goes IN a dispatch prompt (build-workflow.md governs whether to dispatch): lane boundaries, control independence on the graded field only, free self-checks, resumable manifests. Derived from the 130-agent 2026-07-19 run.
- [independent-review-modes.md](docs/rules/independent-review-modes.md) — Blind Review vs Targeted Verification for outside reviewers (Codex); log.md poisons a cold audit; two passes then go empirical.
- [wiki-judgment-norms.md](docs/rules/wiki-judgment-norms.md) — Opus-proofing: the five wiki norms that degrade off-Fable (arc-resolution logging, repeat-surfacing, surface-don't-police, no key-hunts, propose-don't-delete) with wrong-vs-right examples. SPINE §5 points here.
- [business-ops.md](docs/rules/business-ops.md) — open non-project business to-dos (Acoustic, beer-vendor LOW tier, Ott→Burlotto, home page).
- [client-book.md](docs/rules/client-book.md) — **NEW 2026-09-18.** The home-grown CRM for the ~148 whales: the no-bought-seat ruling, what already exists vs. the thin missing layer, `customer_interests` (what a person WANTS), how a terminal lands a want from chat, the 5 slices (memory ✅ → Klaviyo whale sync → `who_wants()` → Fawkes screen → TaT follow), and the Klaviyo 1:1 text route.
- [email-sms-klaviyo.md](docs/rules/email-sms-klaviyo.md) — **NEW 2026-06-20.** Email/SMS platform decision: **Acoustic OUT, Klaviyo IN** (Shopify-native, fed by Supabase/TAT). Two-lane architecture (Klaviyo = marketing/lifecycle, TAT = 1:1/allocation), live pricing (active-profile model, ~$40–70/mo at Thief scale), CAN-SPAM/TCPA consent framework, and the load-bearing **alcohol SMS compliance** rules (DOB age-gate mandatory, form-only opt-in — NOT checkout box; naming products in-body = open Q to confirm w/ Klaviyo). Read before any email/SMS platform, consent-capture, or list-hygiene work.

- [house-registry.md](docs/rules/house-registry.md) — **NEW 2026-09-22, v1.** Which house owns what: every load-bearing thing (code, DB, invoices, mail, board, marketing, Shopify, hosting, DNS, boxes, personal surfaces) → its house, why, blast radius, backup y/n; `⚠ Needs a home` list with proposed placements (Kyle rules) + open questions.
- [expiry-calendar.md](docs/rules/expiry-calendar.md) - **NEW 2026-10-11.** Every dated cliff (API credit, plan switch, batch key, legacy Supabase keys, Fawkes TestFlight, domains, signing certs, Apple membership, lease notices) soonest first, each with its source, owner, blast radius and a mechanical check; born from the 2026-10-10 fringe audit ask #8.
- [dns-subdomains.md](docs/rules/dns-subdomains.md) — **NEW 2026-06-25.** DNS & subdomain registry for `thiefshop.com` (dig-confirmed): every record + who owns it (Shopify apex, M365 MX, Klaviyo `send.` NS-delegated = marketing, Resend `tx.` records-in-GoDaddy = transactional, `_dmarc` parked p=none). **Golden rule: one sending service per subdomain** — born from the `send.` collision (Resend + Klaviyo both on it). Read before any DNS / new-email-service / sending-subdomain work, and the add-a-subdomain checklist before touching GoDaddy.
- [shopify-api.md](docs/rules/shopify-api.md) — Shopify auth (`client_credentials`), location IDs, API gotchas, scopes, `shopify_products` RLS history. **Read before any Shopify code.**
- [shopify-capability-watch.md](docs/rules/shopify-capability-watch.md) — thin watch-list of desired-but-missing Shopify capabilities + load-bearing "can'ts". Verify a "can't" against live docs before trusting it.
- [golden-goose.md](docs/rules/golden-goose.md) — invoice pipeline, daily menu, trust rebuild sessions, Phase 6/8 work.
- [ios-apps.md](docs/rules/ios-apps.md) — Fawkes, TheTell, ThickAsThieves, The Cave, Rail. Swift/iOS conventions.
- [ARCHITECTURE.md](project-fawkes/ARCHITECTURE.md) — **NEW 2026-10-05. The Fawkes code map.** GENERATED from `project-fawkes/docs/architecture/{comms,rail,stock,receiving}.md` + `_header`/`_findings` by `scripts/stitch-architecture.mjs`; `scripts/verify-architecture-map.mjs` greps the Swift/TS source for every RPC · edge fn · table · View struct · deployed fn · POS extension (157 at last run) and fails if any is absent. Per surface: Job · Who sees it · Entry · Screens+gestures (file:line) · Writes · Reads · Downstream · Siblings · State of record. Part 5 = survey findings (rotation landmines, silent writes, dead code, live-only SQL bodies). Read BEFORE touching any Fawkes surface.
- [REDUNDANCY_2026-10.md](project-fawkes/docs/REDUNDANCY_2026-10.md) — **NEW 2026-10-05. The Fawkes redundancy ruling sheet** (closes the "Jackson Pollock" board card). 30-day rows per tab (refreshed from the 09-17 census that was never ruled), same-job doors, dead code, 8 questions with my call pre-filled; Kyle's rulings go under `## Rulings`, each yes → an Approved card + briefed Opus lane. Headline: the "owed to a person" job has 4 doors and 3 wrote 0 rows in 30d — a training gap (Pull was never introduced to staff), not a design verdict.
- [floor-map/ARCHITECTURE-SPOTS.md](project-fawkes/docs/floor-map/ARCHITECTURE-SPOTS.md) — **NEW 2026-11-09, back room added 11-10.** The floor map's data layer as applied: `store_spots` and friends, how a wine's floor and back spot resolve, how to change the map.
- [floor-map/PLAN.md](project-fawkes/docs/floor-map/PLAN.md) + [README.md](project-fawkes/docs/floor-map/README.md) — **NEW 2026-11-04. Fawkes floor + backstock map** ("where is it / where does it go" for day-one staff). README = Kyle's floor rulings from his sketch; PLAN = phases, the count-backstock-only data call, decisions pre-filled. Map trace `ww-floor-map.html`, classifier `spot-default.sql` (94.6% auto-spotted).
- [fawkes-field-guide.html](docs/sops/fawkes-field-guide.html) — **NEW 2026-10-05, v0.** Staff training for Fawkes, by job: 8 printable pages in the `cards.html` pipeline, taps quoted from `ARCHITECTURE.md`, one owner per page. Screenshots pending the iPad walk.
- `docs/rules/the-tell-ai.md` — The Tell AI routing table (Phase B exit, 2026-10-02): tiers, min-tier per feature, measured laws, leak filter, platform facts.
- [subagent-dialing.md](docs/rules/subagent-dialing.md) — **NEW 2026-07-28, PROPOSED not ratified.** Read before ANY fan-out. The three rails (Agent tool / `claude -p` / Workflow `agent()`) and which dials each actually exposes — the Agent tool has **no effort param**, `-p` and Workflow both do. Measured cost from the drift census, the batching-beats-downgrading law, the `ANTHROPIC_API_KEY`-absent check that keeps `-p` on Max quota, and a pre-flight block to paste before firing.
- [react-native-toolchain.md](docs/rules/react-native-toolchain.md) — **NEW 2026-07-28.** The RN/Expo sprint's two non-code blockers, both found by probe: `create-expo-app` cannot scaffold on Node 26 / npm 12 (npm changed `pack --dry-run` from array to keyed object; fix = pinned LTS Node per repo, ⚠ never a global npm downgrade), and the Android half of the toolchain (SDK + JDK) is absent from the Mac — attended install, on the critical path with Play org + D-U-N-S. Xcode is fine, so iOS-side RN via Expo Go works today.
- [rpc-grant-audit.md](docs/rules/rpc-grant-audit.md) — **read before granting or writing any `SECURITY DEFINER` RPC.** The law (never take an authorization input as an argument; PUBLIC gets EXECUTE by default), the two holes closed 2026-07-09, and the proposed-not-run tail awaiting Kyle's ratification.
- [wine-atlas.md](docs/rules/wine-atlas.md) — interactive wine region map. WA + Loire live.
- [enrichment.md](docs/rules/enrichment.md) — thief-wd-rescue scripts, beer/spirits collections, ScraperAPI + wine.com gotchas. **Live corpus↔Shopify coverage figure + the exact denominator lives here** (read it with `scripts/_wave-coverage.mjs`, don't quote an older %). Producer-scrape wave history + the per-wave banked negatives (which producers NOT to re-recon, and why) hang off it in `thief-wd-rescue/_enrichment-spike/<date>-wave<N>/` — wave 8's is `2026-08-22-wave8/banked-negatives.md`.
- [producer-canonicalization.md](docs/rules/producer-canonicalization.md) — vendor/producer name canonicalization across the 3 sources. De-pollution shipped 2026-05-31; full canonicalization PARKED.
- [surface-layer.md](docs/rules/surface-layer.md) — **PARTIALLY SHIPPED 2026-06-08.** The product-page "wrapper": theme `product-details-tags.liquid` now renders Professional Ratings cards (`custom.scores`) + Winemaker/Tasting Notes + producer bio + Details, each self-guarding (`if data != blank`) → mined scores/notes go LIVE when pushed. Still pending: `scores_json` rich notes, Accolade/ABV (dormant — metafields not pushed), body_html migration, producer page; images its own lane.
- [spirits-bundles.md](docs/rules/spirits-bundles.md) — **kill-the-bundle build + migration map.** `wslt-fee` POS extension (built, register-tested, then REMOVED from the deployed app 2026-06-12 + shelved at `_shelved/wslt-fee/` when the kill was stood down), the 2 bugs found+fixed, live footprint (135 standalone spirits), `bundle_parent_id` map, `is_bundle_child` reconciliation, moved-location POS OOS bug + re-wire test plan. **Mechanism is now an open decision** → see `wslt-injection-mechanism.md`.
- [wslt-injection-mechanism.md](docs/rules/wslt-injection-mechanism.md) — **Mechanism decision writeup (2026-06-12).** Applying WA's fixed per-liter spirits tax AUTOMATICALLY (manual = disqualified). The reframe: online can't drop the bundle on Grow (Cart Transform = Plus-only), so "kill the bundle" was only ever POS-deep. Options weighed: keep+fix the bundle (evaluate first), redesigned no-churn reactive POS injection, Plus Cart Transform. Read before any more WSLT mechanism work.
- [shopify-theme.md](docs/rules/shopify-theme.md) — Vantage theme CHANGELOG (what shipped, the bug families). The map is `WebsiteModification/THEME_MAP.md`.
- [label-printing.md](docs/rules/label-printing.md) — Zebra ZD421 label printing. ZPL layout, barcode strategy, priority queue.
- [vestaboard.md](docs/rules/vestaboard.md) — board automation. Edge fn v24, queue table, **pg_cron** schedule, three-lane layout engine (auto/waterfall/verbatim), 6 dynamic content plugs. Reworked 2026-07-18.
- [xero-accounting.md](docs/rules/xero-accounting.md) — Xero apps, Shopify Capital loans, GAAP reminders.
- [store-analytics.md](docs/rules/store-analytics.md) — shelf optimization, WD + Shopify splice plan, gap list framework.
- [shopify-tag-metafield-cleanup.md](docs/rules/shopify-tag-metafield-cleanup.md) — multi-session tag cleanup. Phase A pure-removal ready.
- [profitbuilderhd-migration.md](docs/rules/profitbuilderhd-migration.md) — Kyle's personal WP site → Cloudflare Pages. **Deadline 2026-05-24.**
- [distributor-portal-apis.md](docs/rules/distributor-portal-apis.md) — SGWS Proof + Provi API map. Auth, endpoints, pricing payloads, **Provi enumeration recipe + 28-column schema + parser gotchas** (2026-05-21 update).
- [reorder-system.md](docs/rules/reorder-system.md) — **BUILT AND LIVE** (design of record; read the 07-18 + 07-19 blocks, they win all conflicts). Engine in-database (`run_reorder_engine`, hourly, monitored) + `distributor_schedule`/`distributor_alias`/`purchase_channel` + per-SKU par hook. **07-19: push rail + eyes-only Reorder tab.** `register_fawkes_device` → `fawkes-send-push` (APNs proven live on Kyle's iPhone) → 7:35am day-of digest + hourly event push; `reorder-feed`/`reorder-act` enforce `role='admin'` server-side (staff→403 proven, incl. a staff JWT reading the table directly → `[]`). Tab v2: Wine/Spirits/Beer/Other bucketing, collapsible sections, retire/supersede (local-only, no Shopify writes), On-order, configurable sales window **with horizon detection — real sales history starts 2026-02-01, so any window past ~169d is blind**. Restock hunt Phase 0 + UI also here. ⚠ The email digest hedge is DEAD (superseded 07-18).
- [gg-upload-enrichment-refactor.md](docs/rules/gg-upload-enrichment-refactor.md) — **TRANSIENT handoff brief** (2026-05-25): next session's scope — GG upload idempotency/dedup + created_at hygiene + guarded sync deletion-marking + durable spirits parent-targeting. Delete when shipped.
- [build-vs-buy-review.md](docs/rules/build-vs-buy-review.md) — **Focused dependency review (2026-06-01).** Retry/queue/scheduling plumbing across repos. Verdicts: adopt `p-retry` opportunistically, **NO pg-boss** (serverless mismatch), drop dead `node-cron`. Surfaces 6 retry-gap action items incl. 🔴 shopify-xero-sync has zero retry + 🟠 COGS silent-swallow in `daily-sync.js`.
- [metered-ai.md](docs/rules/metered-ai.md) — hard preflight gate for metered AI/API spend, `claude -p` wrappers, paid third-party services, prod/customer-data boundaries, bulk writes, deploys, and scheduled jobs.
- [dependency-upgrades.md](docs/rules/dependency-upgrades.md) — **Policy of record 2026-07-14.** Dependency/runtime upgrade cadence: capability tiers (T1 writes-external / T2 deployed-read-only / T3 verified-local-only; unclassified ≠ T3), quarterly sweep rides `/retro`, Dependabot alerts-only on 7 repos, majors solo+justified, runtimes separate, npm≥12 install-script gate, lockfiles mandatory in deployed repos, never probe prod token rails from previews.
- [dreaming.md](docs/rules/dreaming.md) — **NEW 2026-07-07. DESIGN locked, skill not yet built.** Out-of-band memory consolidation (the DevCon "dreaming" idea): scheduled headless `-p` pass reads recent session transcripts, finds cross-session recurring mistakes no single window can see, **proposes** wiki/memory fixes (cited + prevalence) for Kyle to ratify. Fills the transcripts×grow quadrant that `/retro` (quarterly-strategic) + `groundskeeper` (wiki-prune) leave empty; feeds them, doesn't replace them. Local launchd (data is on-disk), Opus-never-Fable, propose-only, daily-for-tuning→weekly. Holds the `-p`-goes-metered tripwire.
- [xero-reconciliation-spec.md](docs/rules/xero-reconciliation-spec.md) — **SPEC, brainstorm-next-session (2026-06-01).** The reconciliation "sweep" tripwire for shopify-xero-sync: reconcile by `updated_at` (no window edge), day-granularity recompute, closed-period policy call, two tripwires (invoice AmountDue + Shopify Clearing 1090 drift). Pairs with the shipped `currentQuantity` fix.
- [thief-wd-rescue/xero-backfill/README.md](thief-wd-rescue/xero-backfill/README.md) — **NEW 2026-07-17. Lane E dry-run backfill generator** (frozen WD archive → daily-grain Xero invoice/COGS CSVs; zero posting by construction): money model, resolved decisions (WSLT→2320 · box B re-ruled = payments direct to Banner, no clearing account), and the **deposit-lag model** (Thu+F+S→Mon lump, 2-business-day, holidays separate — the deposit-tie-out spec). Verdict + evidence: `xero-backfill/FABLE_REVIEW_2026-07-17.md` (APPROVED). Artifacts: `xero-backfill/dry-run-<date>/`. Open remnants: `xero-backfill/SHUNTED_DAYS_RULING_PACK.md` (**2026-07-19 — the 3 remaining 2025 item-gap days pre-dug to named orders**: Apr-11 half-refunded return · May-3 zero-dollar exchange · Dec-31 = 41 itemless NYE orders; each a 2-min Kyle ruling).
- [simon-scrapes-funnel-teardown.md](docs/rules/simon-scrapes-funnel-teardown.md) — **Competitor/template intel (2026-06-15).** Simon "Scrapes" Coton's AgenticOS/Agentic Academy funnel: same substrate as us (Claude Code + Skool + dashboard-as-talon), pricing ladder $37/$77/$647, free-challenge activation, member-count price ratchet. **Key read: the dashboard is a retention talon at the mid-tier, NOT the acquisition hook** → confirms Command Deck = calling card; the gap is content/top-of-funnel. Serves [[project_consultant-kyle-diversification]] + Command Deck ROADMAP "Sell gates."
- [distributor-scrape/PROVI_PROGRESS.md](distributor-scrape/PROVI_PROGRESS.md) — live scratchpad for Provi scrape build.
- [distributor-scrape/PROVI_CONFLATION_AUDIT.md](distributor-scrape/PROVI_CONFLATION_AUDIT.md) — audit + priority fix plan for the Provi `external_sku` conflation: derived `distributor_wines` keys Provi houses (`vehrs`/`winebow`/`galaxy`/`walden`/`ambrosia`/`wine_trust`/`vin2u`/`grape`) by the meaningless Provi shell ID, not the real house `distributor_code`. Latent (reciprocal accident, not on fire). **Read before touching `build-distributor-wines.mjs` or Provi-house matching.** Raw `distributor_products` PK — leave alone.
- [distributor-scrape/ANW_PROGRESS.md](distributor-scrape/ANW_PROGRESS.md) — American Northwest (Jet Admin) — 9,391 rows shipped.
- [distributor-scrape/GRAPE_PROGRESS.md](distributor-scrape/GRAPE_PROGRESS.md) — Grape Expectations (Airtable Client Portal) — 1,908 main + 121 Bordeaux. Documents the Cookie-Editor injection pattern for any Airtable Client Portal future scrape.

## Golden Goose (`golden-goose/`)

- [briefs/GOOGLE_REHOME_RUNBOOK.md](golden-goose/briefs/GOOGLE_REHOME_RUNBOOK.md) — **NEW 2026-08-01. Kyle's click-by-click for the GCP sitting** (~15 min, $0, needs a real browser — step 4 ends in a file download). Create the no-org `thief-invoices` project → enable Drive + Generative Language APIs → service account (no project roles) → restricted-from-birth Gemini key → share both store folders Editor → ferry both creds to zanzibar → verify, THEN decommission. Also records why the poller needs two credentials rather than one.
- [briefs/GOOGLE_REHOME_BRIEF.md](golden-goose/briefs/GOOGLE_REHOME_BRIEF.md) — **2026-07-27; Drive half BUILT + VERIFIED 2026-08-01** (two store folders, full partition, `delivery_store` stamped from the folder). Remaining: GCP steps 2–3 + 6–7 → see the runbook above. Phase 1: re-home GG's Google infra out of buildprofit-org + personal-Gmail Drive into a standalone thiefshop GCP project (finally lets the Gemini key be restricted); old key retired via SECRETS_LOG at decommission. Phase 2: poller `--once` flag → zanzibar launchd every 15 min (run-job.mjs + registry + watcher), laptop poller habit dies. Out of scope but captured: review-ui-on-zanzibar (own doctrine sitting), GBP ownership transfer, buildprofit shared drive.
- [ALLOCATION_TAG_BRIEF.md](golden-goose/ALLOCATION_TAG_BRIEF.md) — **NEW 2026-07-24. Ready to fire.** Gives the "Allocated" toggle teeth: product `active` but POS-only (Online Store withheld → 404, proven 6/27) + placement written `status='allocated'` so it never hits the Fawkes Deliveries tab, while staying ringable at the register. **Why not Shopify `draft`: draft is unsellable on ALL channels incl. POS.** Fawkes needs zero changes (`get_pending_placements` already filters on status). Enum widen + 3 GG touchpoints; $1-test acceptance gate with a non-allocated control. First use case = Clape invoice 1388543.
- [ARCHITECTURE.md](golden-goose/ARCHITECTURE.md) — **definitive code map** (verified 2026-06-06). "How the program works / where does X come from / blast radius of Y" in one read. Supersedes the two `_archive/` architecture docs. Structural layer regenerable via `/graphify . --update`.
- [IMAGE_HARDWIRE_DESIGN.md](golden-goose/IMAGE_HARDWIRE_DESIGN.md) — **NEW 2026-07-31, design of record.** Distributor-catalog image attach at product-create (tiers 0–3, probe/flag/queue, every measured number: 403 rates, pixel floor, yield). Corpus lake reaches ~1% of GG-era imageless; `distributor_products` is the real supply. Kyle rulings + the CONTESTED bottle-vs-label ordering both recorded here.
- [IMAGE_HARDWIRE_TIER1_REVIEW_2026-07-31.md](golden-goose/IMAGE_HARDWIRE_TIER1_REVIEW_2026-07-31.md) — **NEW 2026-07-31, closed same night.** Fable Tier-1 of the image lane + taxonomy gate (F1 gap-queue index collision, F2 unscoped-code wrong-image, both fixed + replayed) **+ Codex cross-vendor blind addendum** (7 more findings, 6 fixed — corrected-UPC threading, code-contract tightening, probe deadline). The full maker→Fable→Codex review chain for this lane.
- [ROUTING.md](golden-goose/ROUTING.md) — **NEW 2026-06-11.** The *decision graph* companion to ARCHITECTURE.md's call graph: matcher + classifier fork trees with live counts, ranked verified leaks (#1 = producer mis-bind at conf 1.0, bypasses review, already uploaded), killed theories, refresh SQL. Interactive twin: `golden-goose/ROUTING_TREE.html`.
- [DISTRIBUTOR_FIRST_MATCHER_DESIGN.md](golden-goose/DISTRIBUTOR_FIRST_MATCHER_DESIGN.md) — **NEW 2026-06-11. APPROVED build spec.** Distributor-first matcher rework: catalog (`distributor_wines`) as router, LWIN demoted to enrichment. `distributor_links` ledger (merges `product_supplier_codes`, precedence-guarded), bridge maps on `suppliers`, miss = fuzzy + `distributor_catalog_gaps`. Per-distributor coverage table + phased seed (SGWS/ANW first). Builds against ROUTING.md's write-back contract.
- [APPELLATION_CANON_SCOPE.md](golden-goose/APPELLATION_CANON_SCOPE.md) — **the wine-geography canon plan** (born from the 06-29/30 Rhône drift). Canon SOURCE = **GuildSomm Compendium** (richer than `wine_regions`, which demotes to gap-fill) → `guildsomm_compendium_raw` via `scrape-guildsomm-compendium.mjs`. **France crawled ✓ (423 pages, 2026-07-01); parse pass = next build; US = 2nd crawl target.** Holds the canon schema, the mid-bucket/sub-region facet model (grape-can't-split heuristic + geo vs classification flavors + per-region SKU table), the Burgundy 4-axis (village+classification, locked-out crus) + California un-flatten mirror, build order, and GG resolver design.
- [PROJECT_GOLDEN_GOOSE.md](golden-goose/PROJECT_GOLDEN_GOOSE.md) — high-level project description, phase roadmap, vision. (Current *code map* = ARCHITECTURE.md above; this doc's own architecture section is older.)
- [GOLDEN_GOOSE_CURRENT_AUDIT.md](golden-goose/GOLDEN_GOOSE_CURRENT_AUDIT.md) — 2026-05-25 code-first audit; open risks list (atomic claim, recovery sweep, ingest partial-state).
- [INTENT_AUDIT_2026-07.md](golden-goose/INTENT_AUDIT_2026-07.md) — **NEW 2026-07-01 (Fable week), IN PROGRESS.** Intent-vs-reality audit: where code+docs agree with each other but disagree with Kyle. F1 (spirits UPC layer) + F4 (duplicate factory) SHIPPED 7/01–02; F5 mechanisms traced (M3 = staff create-at-POS on dead scans). Read before any GG fix work this month.
- [INTENT_INTERVIEW_2026-07.md](golden-goose/INTENT_INTERVIEW_2026-07.md) — **NEW 2026-07-07. Phase C of the intent audit, COMPLETE.** Kyle's intent captured unanchored, per-lane (beer/wine/spirits + shared tail): trust ladder, self-heal/stone-tablet rule, publish-immediately, constraint-driven matching, allocated-feature gap, queue=parking-lot, staff-QA checklist. Ends with a 10-theme synthesis crib + collected verify-flags. Phase D reads this + the audit file together.
- [INTENT_FINDINGS_2026-07.md](golden-goose/INTENT_FINDINGS_2026-07.md) — **NEW 2026-07-07. Phase D DELIVERABLE — the ranked findings doc, awaiting Kyle's adjudication.** 16 findings (D1–D16) in 5 pain-ranked tiers, each tagged drifted/unfinished/never-built + fix-shape + effort + checkbox. Synthesizes Phases A+B (audit file) × C (interview). Approved boxes become the hardening sprint (B terminal packages briefs). Read this FIRST for "what's wrong with GG and what do we do."
- [F5_CONSOLIDATION_BRIEF.md](golden-goose/F5_CONSOLIDATION_BRIEF.md) — **NEW 2026-07-02. Hand-off brief (Opus).** Consolidate the 25+ duplicate active pairs: census → per-pair survivor plan CSV → Kyle gate → apply w/ read-back. Clement playbook + the week's landmines baked in. Data-only job; pipeline code off-limits.
- [TRUST_REBUILD_SESSIONS.md](golden-goose/TRUST_REBUILD_SESSIONS.md) — 4-session punch list. Sessions 1-4 shipped. **Note: docs may sound worse than current reality — needs reconciliation.**
- [MENU_REFACTOR_PLAN.md](golden-goose/MENU_REFACTOR_PLAN.md) — daily menu refactor (2026-05-09). Phases 1-5b shipped. **Superseded 2026-06-06** by the menu-tool redesign spec (held-in-state→no, photo path + web wrapper retired, Approach A): `docs/superpowers/specs/2026-06-06-menu-tool-redesign-design.md`. Keep as reference for the proven pieces (lookup-v2, bottle-tag, menu_notes).
- [PHASE_6_HANDOFF.md](golden-goose/PHASE_6_HANDOFF.md) — `/query` slash command brief for shop Mac.
- [PLAYBOOK.md](golden-goose/PLAYBOOK.md) — how to run multi-session work without getting lost. Session prompt pattern.
- [UPLOAD_CONTRACT.md](golden-goose/UPLOAD_CONTRACT.md) — required vs best-effort stages per upload action. Invariants.
- [POUR_SYNC_SETUP.md](golden-goose/POUR_SYNC_SETUP.md) — staff computer setup for pour sync tool.
- [RECEIVING_STATION_DESIGN.md](golden-goose/RECEIVING_STATION_DESIGN.md) — **NEW 2026-07-04. Phase 6 scope, Kyle-locked design.** The staff-owned last mile (pallet → checked/labeled/on floor): QA-checklist-as-taps + flags, house-UPC label printing (Zebra ZD421 already owned), New Arrival tag nomination. **Two-UPC "Arizona rule": `shopify_products.legacy_upc` (invoices/matcher) + `barcode` (house UPC, POS) — do not merge, ever.** Rollout = receiving forward-going + section-a-week sweep that doubles as cycle count/dead-scan audit. Supersedes the May UPC-bridge plan; absorbs the Inventory Counting System plan.
- [SPIRITS_UPC_RECOVERY.md](golden-goose/SPIRITS_UPC_RECOVERY.md) — **NEW 2026-06-13. Handoff brief.** ~255–288 spirit bundle parents missing UPC. Forensic conclusion (in the brief — supersedes earlier "wipe" theory): most likely never fully entered — barcode presence correlates strongly with in-stock-at-transfer, the signature of stock-gated hand-scanning, not a bulk wipe. Recovery = match `distributor_products.upc` → write to parent variant (dry-run → Kyle approves → read-back verify). Open Q: parent vs child target (resolve WSLT-on-scan first) — now folded into `INTENT_AUDIT_2026-07.md` F1. Scripts in `review-ui/scripts/`.

## Clock (`timeclock/`) — the Homebase replacement
- [timeclock/README.md](timeclock/README.md) — **NEW 2026-11-15.** The app's front door: what talks to what, layout, the three verifier commands, which rules Kyle ruled vs which are Claude's proposals, and the Homebase parity map (built / in a package / not built), from Homebase's own feature pages.
- [timeclock/gusto/README.md](timeclock/gusto/README.md) — **NEW 2026-11-15.** Gusto's 17-column hours import template (as pulled by Kyle), the template filler in `core/gusto.ts`, what the Gusto account holds (employees, pay schedule, earning types), and how a courier would write through the claude.ai Gusto connector (`update_payroll`; never `record_time` before cutover).

## TAT Android (`TaT-rn/`)

- [TaT-rn/README.md](TaT-rn/README.md) — **NEW 2026-08-03.** The RN/Expo Android client for Thick as Thieves. States the in/out boundary a folder name can't carry (iOS stays SwiftUI in `ThickAsThieves/` and is **never** being converted; the Tell is separate by **ownership**, not tech), why TAT deliberately lives in two codebases and what that implies (**logic belongs in Supabase RPCs, clients stay thin**), plus the gotchas that already cost time: the `.npmrc` `allow-scripts=` fix, the pinned Node, and the stale-Metro trap. Spec of record: `docs/superpowers/specs/2026-08-03-tat-android-v1.md`.

- [TaT-rn/store-listing/README.md](TaT-rn/store-listing/README.md) — **NEW 2026-08-14.** The staged Play production checklist, so the paperwork sitting is review-and-click. Index to `forms.md` (pre-filled Data-safety / content-rating / App-access answers, verified against the client not the iOS packet) and `copy.md` (title, 3 short-description options, full description). **Its one blocker is CLEARED 2026-08-14:** Play requires a WEB account-deletion URL, and `thiefshop.com/pages/delete-account` is now live (Apple never asked, so `APP_STORE_PACKET.md` had no answer to inherit). Page is owned by `WebsiteModification/scripts/create-delete-account-page.mjs` — edit its `BODY` and re-run to change the wording. Also records why raw 1080×2424 Android shots are silently out of Play spec (max dimension capped at 2× the min) and the open gold question `BRAND.md` §2a gets wrong.

- [TaT-rn/store-listing/SUBMIT-A-RELEASE.md](TaT-rn/store-listing/SUBMIT-A-RELEASE.md) — **NEW 2026-09-07.** How to ship a new Android version, end to end: bump `app.json`, build the bundle, **prove the signer** (a debug-signed `.aab` looks perfect on disk and is rejected at upload with nothing useful said), then the Play Console click path. Written while submitting `3 (1.10.0)` because the 08-19 first submission was the only one and the steps lived in nobody's head — Play's model is a **track** you replace, not an Apple-style version you submit, which is most of why it reads as strange. Also names the two traps: a reused `versionCode`, and the reviewer age gate that disarms itself on every sign-in.

- [TaT-rn/PAYMENTS_PARITY_SPEC.md](TaT-rn/PAYMENTS_PARITY_SPEC.md) — **NEW 2026-08-26.** What it takes to bring Android up to iOS 1.9, where "payments" is **three rails with three different gates** (cart = retail only · pay-my-hold = any tier · pay-my-claim = retail only — deliberately not harmonised, and the spec says why). Headline finding: **the server is already done** — all three edge functions are platform-agnostic, Android push is live, `tat_app_config.android_min_build` already exists, so this is a pure client build with no migration. Carries the paid-marker state machine rule-by-rule (each one an adversarial-review finding, on real money), the two RN-vs-Swift kit API mismatches (no `cancelled` event; `invalidate()` is NOT the Shop Pay reset), Lane 0 as a gate before any lane is briefed, and the Play answer that Apple's 3.1.5(a) dance doesn't apply — physical goods must NOT use Play Billing.
- [TaT-rn/docs/PASSPORT_ANDROID_PORT.md](TaT-rn/docs/PASSPORT_ANDROID_PORT.md) — **MERGED onto main 2026-10-12** (was the `passport-android` worktree). The Passport Android port map: one row per iOS Swift file → Android home, the five RPCs, the map-data bundle, P1–P4 done, rulings D1–D15 all ruled by Kyle 10-10 (D5 executed = `expo-sharing`).
- [TaT-rn/docs/DROPS_112_PARITY.md](TaT-rn/docs/DROPS_112_PARITY.md) — **MERGED 2026-10-12.** Drops 1.12 on Android (Buy-first, the released card, 3-day hold copy): iOS change → RN file:line, divergences ruled.
- [TaT-rn/docs/CORKY_ANDROID.md](TaT-rn/docs/CORKY_ANDROID.md) — **NEW 2026-10-12.** Corky on Android: iOS file → RN file:line, the haiku → deterministic ladder (no PCC), the post-check port graded by hash against the shared Node verifier, divergences (incl. the two iOS bugs it found), Codex rounds.
- [TaT-rn/docs/R8_ANDROID.md](TaT-rn/docs/R8_ANDROID.md) — **NEW 2026-10-12.** R8 minify + resource shrink on Android release builds: `plugins/withR8.js` (the live SDK-57 property name, throws if the template changes), keep-rules file, walk coverage, size delta.
- [TaT-rn/docs/TOAST_ANDROID.md](TaT-rn/docs/TOAST_ANDROID.md) — **NEW 2026-10-12.** The iOS Toast port: Swift → RN file:line, rules of use, 7 sites + 3 hosts, offline-from-the-error, divergences, Codex rounds.
- [TaT-rn/docs/OFFERS_ANDROID.md](TaT-rn/docs/OFFERS_ANDROID.md) — **NEW 2026-10-12.** Offers + the offers cart + the "Your picks" review on Android: iOS file → RN file:line, the one `buy-offer` call site, divergences (no Safari fallback, in-memory paid markers), Codex rounds.

## Project Fawkes (`project-fawkes/`)

- [PROJECT_FAWKES.md](PROJECT_FAWKES.md) — vision doc. "How do I stay in the life of someone who was blown away by The Thief after they go home?"
- [project-fawkes/BUILD_STATE.md](project-fawkes/BUILD_STATE.md) — Swift conventions, current build state, model gotchas.
- [project-fawkes/METRICS_SPEC.md](project-fawkes/METRICS_SPEC.md) — **NEW 2026-08-15, rewritten the same night for v2 + v3 (Phases 0–3, each with its controls block at the foot; where the body and the controls disagree, the controls win).** Phase 3 is the rollup layer — the owner deck reads nightly tables in a private `metrics` schema (10.6s → 3.0s) while the `v_metrics_*` views stay the only definitions. Spec of record for the Metrics tab: the four books (WW counter / WW bar / North Side / online), what each metric means and what it deliberately excludes (tax, tips), the per-staff `user_id` back door and the schedule trap that comes with it, the four-direction auth replay, and the open list. Read before touching `metrics-feed` or the `v_metrics_*` views.
- [project-fawkes/METRICS_REVIEW_2026-08-15.html](project-fawkes/METRICS_REVIEW_2026-08-15.html) — **NEW 2026-08-15.** Fable's review of the v1 build + the Kyle sitting: verified holes (pseudo-profiles → honest carded 33–41%; WD/Shopify id seam → new customers 20–39% inflated; tip heuristic deletes custom sales; uncosted glass pours → WW GM 31% on costed lines), what belongs on each screen, the 7-agent web research (raw: `.research.json`), decisions D1–D17 with the live rulings stamped. Read before touching the v2 build.
- [project-fawkes/METRICS_V2_BRIEF.md](project-fawkes/METRICS_V2_BRIEF.md) — **NEW 2026-08-15.** The v2 build brief for one Opus terminal: Phase 0 data fixes (pseudo table, identity-resolved new customers, `line_title`/`is_tip`/`rail_tab_id` backfill, pour costing, `v_metrics_book_month`) → Phase 1 staff scoreboard v2 → Phase 2 owner deck v2 → **Phase 3 rollup layer** (added after Kyle reopened the lane), with controls and [K]/[F] provenance on every decision. Fable reviewed at each phase exit.
- [project-fawkes/DEVICE_TEST_CHECKLIST.md](project-fawkes/DEVICE_TEST_CHECKLIST.md) — **NEW 2026-07-27: the pre-TestFlight device sitting** (scanner = only hardware unknown) + the ship tail in Kyle's revised order (iPads re-login AFTER TestFlight, same-day). Count-gaps ruled ship-without 07-27.
- [project-fawkes/INVENTORY_COUNT_DESIGN.md](project-fawkes/INVENTORY_COUNT_DESIGN.md) — ROUGH DRAFT (Kyle 2026-11-04; superseded for counting by YEAR_END_COUNT_SPEC.md) — orig. PROPOSED 2026-07-16: count sessions (rack + Volcano Lair zones), label-migration interplay, Scan tab v1. Kyle ratifies before build.
- [project-fawkes/REORDER_TAB_BRIEF.md](project-fawkes/REORDER_TAB_BRIEF.md) — **2026-07-18. Executable Opus brief, priority one:** order-day push per distributor + Kyle's-eyes-only Reorder tab (edge-fn gated, floor@ iPad migration = hard gate). **Phase 0 SHIPPED same day** — see `docs/rules/reorder-system.md` 2026-07-18 blocks.
- [project-fawkes/PULL_TAB_BRIEF.md](project-fawkes/PULL_TAB_BRIEF.md) — **2026-08-24. Executed Opus brief + Kyle's 08-25 amendments (§6–8):** one Pull tab = every bottle owed to a named customer (claims · Shopify orders · specials; NO holds section — a hold is created by banding; `rail_hold_id` orders chip as Hold). **SHIPPED 08-25** — fns live, migration awaits Kyle release.
- [project-fawkes/RESTOCK_HUNT_BRIEF.md](project-fawkes/RESTOCK_HUNT_BRIEF.md) — **NEW 2026-07-18. Executable Opus brief, ~1 day:** Restock pin / save-to-hunt / not-found / **ghost-bottle** detection + undo. Ghosts are invisible to the reorder engine, so this is its missing sensor. **Must ship in the same Xcode sitting as the reorder push entitlement** (Restock = ~76 staff actions/day).
- [go-thiefshop/README.md](go-thiefshop/README.md) — QR redirect layer for bottle labels (`go.thiefshop.com/p/[handle]`). Vercel; DNS pending in GoDaddy.
- [project-fawkes/FABLE_TIER1_REVIEW_2026-07-19.md](project-fawkes/FABLE_TIER1_REVIEW_2026-07-19.md) — **Tier-1 review of the reorder+restock build. Verdict ship-with-fixes.** Found C1: staff could self-promote to admin and defeat the eyes-only law (proven, fixed, verified) + H1 restock-feed unauth + M1/M2. Pairs with the `LESSONS_LEARNED.md` "Revoke That Revoked Nothing" entry.
- [project-fawkes/FABLE_TIER1_REVIEW_2026-07-19-PASS2.md](project-fawkes/FABLE_TIER1_REVIEW_2026-07-19-PASS2.md) — second pass over the C1 fix + the M1/hunt-UI/reorder-v2 work that landed after pass 1.
- [project-fawkes/FABLE_TIER1_REVIEW_2026-07-19-PASS2-VERIFY.md](project-fawkes/FABLE_TIER1_REVIEW_2026-07-19-PASS2-VERIFY.md) — **third context, catalog-sourced: independently confirms the C1 escalation is CLOSED** (attacked `role` via grant surface, every SECDEF *and* invoker fn, the signup-trigger path, and RLS). Corrected the maker's reading of his own bug — `information_schema.column_privileges` projects TABLE grants onto every column, so a table grant reads as a column grant there. F1 (finish the revoke) and F3 (`is_shared_login` flag replacing an email-prefix match) **done 07-19**; roster-behind-auth + `floor@` password rotation re-timed to "before a shared floor login exists" (Lanes card).
- [project-fawkes/CODEBASE_AUDIT_2026-04-19.md](project-fawkes/CODEBASE_AUDIT_2026-04-19.md) — security + reliability audit. Most items shipped in Build 12.
- [project-fawkes/SHOPIFY_POS_EXTENSION.md](project-fawkes/SHOPIFY_POS_EXTENSION.md) — The Rail POS handoff research + setup.
- [project-fawkes/pos-extension/extensions/players-card/README.md](project-fawkes/pos-extension/extensions/players-card/README.md) — Players Card POS scan: design, wire contract, deploy/test steps, secret rotation, follow-ups (built 2026-07-18). **v1.2 mailing-address capture staged 2026-08-21 — cohort gate + its ceiling, the save contract, and Kyle's deploy/register-test steps live here.**
- [project-fawkes/FIGMA_MOCKUPS.md](project-fawkes/FIGMA_MOCKUPS.md) — Phase 2 design specs.
- [FAWKES_KPI_DASHBOARD_PROMPT.md](FAWKES_KPI_DASHBOARD_PROMPT.md) — metrics/KPI tab brief for next session.

## Wine Club + Blind Tasting

- [PROJECT_WINE_CLUB.md](PROJECT_WINE_CLUB.md) — The Valley / The Route / Blind Box. Sprint 1+2 infra done. Edition 1 next.
- [Blind-Tasting-App/Project_BlindTasting.md](Blind-Tasting-App/Project_BlindTasting.md) — fingerprint engine, 62 fingerprints, decision trees.
- [Blind-Tasting-App/CORPUS_AUDIT.md](Blind-Tasting-App/CORPUS_AUDIT.md) — Layer 1+2 audit output (2026-06-02): 365 findings across all 62 fingerprints, sorted CRIT→LOW with exact evidence + fix direction. Dominant defect = inverted `never_rules` (knowledge correct, logic column sign-flipped). Read before any corpus-fix work.
- [Blind-Tasting-App/CURRICULUM_GUIDE.md](Blind-Tasting-App/CURRICULUM_GUIDE.md) — vintage typicality 2004-2024, producer testability flags.
- [db-migrations/2026-09-16-0001-passport-madeira-coordinate/manifest.md](db-migrations/2026-09-16-0001-passport-madeira-coordinate/manifest.md) — **⛔ SUPERSEDED 2026-09-16, never run — do not release.** Wrong lane (`passport_places` is Rebuild: DML via `execute_sql`, no package); Madeira landed via `curated/pins.json` + the rebuild-lane sync instead, and the preimage gate now refuses on its own. Kept as a record. Originally: one row: `passport_places` puts Madeira **1,129 km** out in mainland Portugal (a Wikidata mis-resolve); `places.csv` was corrected 2026-09-15 from the EU-PDO centroid for PDO-PT-A0038 and this carries it into the table. No DDL, no grants, exact rollback. Nothing reads `lat`/`lng` today — checked in the DB (`pg_proc`/`pg_depend`) and in the app (the map projects from the already-corrected CSV) — so it is **drift, not a bug**, filed as a correction. Preimage gate rehearsed read-only against production: it opens. Read-back carries a positive control (Barolo's pin) so an empty result cannot read as quiet passes.
- [db-migrations/2026-09-13-0003-passport-wine-only/manifest.md](db-migrations/2026-09-13-0003-passport-wine-only/manifest.md) — **RELEASED 2026-09-13 evening** (Codex R1 + R2 taken, readback green). Kyle's ruling: the Passport is WINE ONLY. Both RPCs filter `is_wine` unconditionally; the client flag is ignored. Empirical borrowed-customer gate with an old-RPC baseline. Cost at identity grain: 993 stamps / 890 identities; 2 bundle-only TaT accounts reach the empty state (never held a stamp). Follow-up: drop the fossil `p_wine_only` once the next TaT build ships.
- [db-migrations/2026-09-13-0001-cellar-passport-separation/manifest.md](db-migrations/2026-09-13-0001-cellar-passport-separation/manifest.md) — **NEW 2026-09-13, REHEARSED GREEN on final bytes (+ round trip), Codex ×2 SHIP WITH FIXES (all taken) — RELEASED 2026-09-13, readback green.** Kyle's ruling 16: the TaT Cellar and the Passport are two children of one parent (`customer_wine_history`) and share nothing below it. The 09-10 package had extracted the Cellar's bottle CASE into `passport_line_class` and made both read it ("so they cannot drift apart" — one clause, never put to Kyle); 0012's passport tightening would then have removed 2,057 packaged-beer titles from every Cellar. This package gives each its own function — `cellar_line_class` (byte-identical, md5-pinned) and `passport_stampable` (v1 = behaviour-identical) — re-points the RPC by one token and the view by one WHERE, and DROPS the shared organ. Controls: the view's keyed row set identical both ways (183,305 rows, 0 gone / 0 appeared); all 16,336 distinct till lines classify identically; RPC md5 with the name swapped back = preimage. Generator: `ThickAsThieves/passport/generate-migration-split.mjs`.
- [db-migrations/2026-09-13-0002-passport-wine-fixes/manifest.md](db-migrations/2026-09-13-0002-passport-wine-fixes/manifest.md) — **NEW 2026-09-13, REHEARSED GREEN on final bytes (+ round trip), Codex ×2 (R1 HOLD → all taken; R2 SHIP WITH FIXES → all taken) — RELEASED 2026-09-13 after 0001, readback green: 95.3% coverage, 5,044 residue lines.** The passport-only half of 0012: the home-state alias fix (7 rows), `passport_stampable` v2 (Kyle's 09-12 rulings — pours, flights, sake/aperitivo/cider/NA, fees), the twin table rebuilt with the evidence tie-break (ruling 15). The view is NOT replaced (0001 already re-pointed it → no AccessExclusiveLock window). Cannot reach the Cellar by construction — both Cellar md5s asserted unchanged. Rehearsed: wine lines 125,105 → 106,834, placed 106,999 → 101,790, residue 18,106 → 5,044 (denominator), non-wine 58,200 → 54,975 (packaged beer stays, asserted). **The movement control fired on its first run — 169 lines moved, not 129** — and the 40 extra were the twin table catching up with two tags Kyle fixed that morning (Guigal → Côte-Rôtie, Lafon Santenots → Volnay); transitions are now CLASSIFIED (named / catalog-led / unexpected → abort). **Rehearsing the rollback found 0012's rollback could never have run** (named a GENERATED ALWAYS `id`). Residuals: canon row 8238 still wrong (Lane B), twin table is a snapshot with no rebuild cron, passport-is-wine-only is package 0013.
- [db-migrations/2026-09-12-0012-passport-lane-a/manifest.md](db-migrations/2026-09-12-0012-passport-lane-a/manifest.md) — **SUPERSEDED 2026-09-13, NEVER RELEASED — do not apply.** Split into the two packages above. Kept as provenance: its survey (the Columbia Gorge alias defect, whole-canon sweep), `evidence/codex-round-1.md` (13 findings on the same rules, incl. the `[0-9]+\s*oz` P1 that matched the whole packaged-beer catalog), and the rehearsal that found the alphabetical tie-break. Carry regardless: `statement_timeout` is 2 min PER STATEMENT and one scan of `v_passport_placements` costs ~19s; 74.6% of placements come from the FROZEN twin snapshot whose rung outranks the live ones.
- [db-migrations/2026-09-11-0004-claim-modify-guards/manifest.md](db-migrations/2026-09-11-0004-claim-modify-guards/manifest.md) — **NEW 2026-09-11, STAGED — needs Kyle's RULING (`RULING.md`) before the release.** Closes four holes in `claim_drop`'s modify path, three carded out of scope by `2026-09-04-0004` and a fourth found in review: a **paid** claim was resizable both directions; a customer could resize **their own allocation** (the guard asked `claim_rule='fcfs'` and **41 of 41 drops are `fcfs`** — `drop_type` is what discriminates); an **increase** never re-ran `tat_account_matches_drop`; and a **collected** claim was resizable (`claim_picked_up` needs no banding mark). All four reproduced on the LIVE body before any fix. Server-only — both clients already render an unknown code's server `message`. Codex round 1: request changes, nine findings, two P1 — **one corrected the package's own evidence** (the checkout HMAC IS re-verified at settlement by `collect-holds-on-order`; the undefended window is *after* `paid_at` is stamped). Named residual, not done: the webhook's final write re-checks status and `paid_at` but **not quantity** — a Fawkes edge-function change. Exposure today: 0 of 4 claims paid or collected, 0 allocations open.
- [db-migrations/2026-09-11-0003-event-push-audience/manifest.md](db-migrations/2026-09-11-0003-event-push-audience/manifest.md) — **NEW 2026-09-11, RELEASED 2026-09-11** (both RPCs verified live in `pg_proc` 2026-09-11 evening; the "STAGED" this line first carried was stale within hours). The events push rail moved server-side: `get_event_push_audience(p_platform)` + `tat_event_push_budget()` (both SECURITY DEFINER, service_role only — they return device tokens), `tat_accounts.is_internal`, two `tat_push_policy` columns, and a partial unique index on `(account_id, wave_key)` that makes a keyed send's INSERT an **atomic claim**. Pairs with `send-push` EVENT MODE. Audience control was an **equality: 81 = 81** vs the predicate the two box scripts used; the 21+ gate excludes 0 today. `enforce_event_budget` ships FALSE (quiet hours are enforced regardless). Two Codex rounds, fifteen findings, five P1s — dispositions in `evidence/codex-round-{1,2}.md`, including the three ruled-not-fixed with reasons. ⚠ Release order matters and the reverse order breaks the Thursday cron *after* the SMS has gone.
- [db-migrations/the-tell/2026-08-23-0002-scope-profile-writes/manifest.md](db-migrations/the-tell/2026-08-23-0002-scope-profile-writes/manifest.md) — The-Tell migration 0002: scopes `daily_tell_profiles` writes to the player-safe columns. Closes the **third** instance of the additive-column-privilege law (a column REVOKE under a table GRANT is a no-op). Exploit replayed both directions; first The-Tell migration through the guarded MCP rail.
- [Blind-Tasting-App/tools/daily-tell/README.md](Blind-Tasting-App/tools/daily-tell/README.md) — **The Daily Tell**: Wordle-shaped daily blind-call game on the engine (62-dot board, 6 tells, shelf-bottle reveals). Regen commands + honesty contract inside. Built 2026-07-17; fork opened 2026-07-22.
- [Blind-Tasting-App/tools/daily-tell/ROLLOUT_BRIEF.md](Blind-Tasting-App/tools/daily-tell/ROLLOUT_BRIEF.md) — public-launch brief: web-first, Supabase anonymous+OTP accounts, plan pinning, live shelf bottle. **LIVE 2026-08-24** at **https://thetell.wine** (www 301s to it; `daily-tell.pages.dev` is the same project, not a second site). ⚠ Two Kyle reversals recorded in it: **Briefing opt-in OUT of v1** (box 3 — no player identity crosses to Klaviyo/the shop) and **shop attribution stripped from the page** (§7). Acceptance tests 1-4 PASS live; 9 is half-done (cron registered + launchd-proven, **alarm never heard**) and 10 open. Landmines section carries the OTP-length bug that made the claim flow 100% dead. Tell iOS webview = Phase 2, TAT parked.
- [Blind-Tasting-App/tools/daily-tell/KYLE_SITTING.md](Blind-Tasting-App/tools/daily-tell/KYLE_SITTING.md) — the launch sitting runbook, **steps 1-6 DONE 2026-08-23** (domain · free Supabase org · auth toggles + Manual Linking verified ON · migration · keys · Resend SMTP on `tx.thetell.wine`) + 6b OTP length ruled 6. All 9 steps closed 2026-08-24. Open: dead-man alarm unproven, test 10. OTP mail copy (Daily-Tell-signed, no Thief line): `otp-email-templates.md`. Live provers beside it: `probe-live.mjs` (RLS both directions, real JWTs) and `claim-flow-live.mjs` (Manual Linking + claim-is-a-conversion; sends a real OTP).
- [TheTell/the-tell-apple-intelligence-spec-v1.3.1.md](TheTell/the-tell-apple-intelligence-spec-v1.3.1.md) — Apple Intelligence project spec (ARCHITECTURE FINAL, 4 red-team rounds): certify → benchmark → build, three-tier model routing.
- [TheTell/PHASE_B_BRINGUP.md](TheTell/PHASE_B_BRINGUP.md) — Phase B environment checklist + live state: 3 lanes done (sim probe PASSED 08-07 — on-device model runs in iOS 27 sim, no physical iPhone needed); gate OPEN (PR-6 executed 07-29, Phase A complete). Remaining Apple-side: AI waitlist (blocks Tier-2/PCC) + SBP activation.
- [TheTell/benchmark/README.md](TheTell/benchmark/README.md) — the §4 benchmark harness (built 08-07): sim-side FoundationModels probe + Node runner + six §4 suites + certified-suite generator off the eligibility view. Run commands, tier coverage, first findings (zero guardrail refusals on tasting vocab; Tier-1 leaks answers in coaching prompts). $0, local, JSONL scoreboard in `runs/`.
- [TheTell/benchmark/corky/](TheTell/benchmark/corky/build-fixture.mjs) — the Corky (TaT somm) probe: `build-fixture.mjs` (5 real Cellars × 3 shapes, no names) · `-CorkyProbe` in The Tell (borrowed PCC seam) · `run-haiku.mjs` · `grade.mjs` → `grade.html`. Design: `ThickAsThieves/docs/CORKY.md`.
- `TheTell/benchmark/TIER2_VERDICT_2026-09-02.md` — Tier-2 (Apple PCC) verdict on the certified suites: Tier-1 vs Tier-2 table, Fable's judged grades, the 11-row Kyle spot-check list, fixture bugs with pre-filled calls.
- `TheTell/benchmark/judge/pcc/CODEX_REVIEW_2026-09-02.md` — Codex blind methods review of the Tier-2 verdict (8 ranked findings; reconciliation lives in the verdict's §Codex review).
- [TheTell/SPEC_GROUND_TRUTH.md](TheTell/SPEC_GROUND_TRUTH.md) — spec v1.3.1 verified against live stack (2026-07-04): engine = RPC not Swift, never_rules is dark data, PR-1 RPC trap, Blocker #0. Read WITH the spec.
- [TheTell/grid-spec/GRID_SPEC.md](TheTell/grid-spec/GRID_SPEC.md) — **Grid Mode design of record (2026-08-14, executed same-session):** prose→scale normalization (0 residue, AXES in `normalize-attributes.mjs` = single source of truth), engine v2 structural scoring (HARD/STRONG/SOFT + 1-step tolerance), deterministic reveal teach, scoped `get_reveal_content` read path. Ceilings named. Migrations of record: `grid-spec/migrations/MIGRATIONS.md`.
- [TheTell/ADJUDICATION_2026-07-04.md](TheTell/ADJUDICATION_2026-07-04.md) — the 6 Phase-A1 calls (Syrah→California, Rioja rescope, MLF policy…) made by Fable at Kyle's request; ratify boxes pending.
- [TheTell/DISPATCH_PACK.md](TheTell/DISPATCH_PACK.md) — self-contained PR-1→PR-6 briefs for post-Fable Opus sessions: acceptance tests, landmines, "what a cheaper model gets wrong" per PR.
- _PR-4 pre-graded inputs (Fable, 2026-07-04 + 07-07 — ALL grading now complete): `TheTell/never-rules-pregrade.csv` (335/335 rules) · `TheTell/elimination-power-scan.csv` (71 always+hard) · `TheTell/never-hard-links-scan.csv` (224 never+hard: 94 keep/122 demote/8 flag) · `TheTell/strong-tier-collapse.csv` (115: 13 hard/99 soft/3 flag) · **`TheTell/pr4-kyle-review.csv` (Kyle's 41-row queue)**. Rubric: `TheTell/pr4-grading/RUBRIC.md`._
- [TheTell/curate-ux-teardown.md](TheTell/curate-ux-teardown.md) — competitive teardown of curate.wine (2026-07-11): $29–49/mo CMS/WSET note builder, agency-grade UX (Söhne, bespoke system), NO reasoning engine. Steal list (chips/rails/negatives/confirm-the-nose funnel), capture-schema implications (read before the v1 capture ratify), Mistake Map → "Tell Map" marketing idea, perpetual font-licensing math.
- [TheTell/MORNING_BRIEF_2026-07-07.md](TheTell/MORNING_BRIEF_2026-07-07.md) — overnight night-shift brief: the 2 Kyle decisions (PR-1 slug ratify box, PR-4 41-row review queue) + everything the night produced.
- [TheTell/migrations/PR1_APPLY_RUNBOOK.md](TheTell/migrations/PR1_APPLY_RUNBOOK.md) — PR-1 apply gate: one ratify box (slug partial-unique deviation) → one command + verification battery + rollback. Migration SQL + 13/13 rehearsal evidence live beside it.
- [TheTell/migrations/PR2_APPLY_RUNBOOK.md](TheTell/migrations/PR2_APPLY_RUNBOOK.md) — PR-2 apply record: **✅ APPLIED 2026-07-15 Kyle-present** (MLF truth layer on all 62 + 4 rescopes as v2 ids 120-123; battery + Grid 62/673/260/237 green; Madeira nod + legacy-crus scrub cleared at the sitting). Rollback valid until a downstream PR builds on ids 120-123.
- [TheTell/migrations/PR4_APPLY_RUNBOOK.md](TheTell/migrations/PR4_APPLY_RUNBOOK.md) — PR-4 (never-rules typing + elimination-power reconciliation): rehearsed green on mig4 clone; **rulings folded per `pr4-flag-rulings.csv`** (both-twins + 11 flags); awaits re-rehearsal → Fable Tier-1 → Kyle-present apply.
- [TheTell/pr3-recon/RECON_NOTES.md](TheTell/pr3-recon/RECON_NOTES.md) — PR-3 read-only recon: the ~90% lookalike-match claim is false (176 external-stub candidates → re-scope with Kyle), attribute-key census (81), negatives split, style two-layer confirmation.
- [TheTell/files/ISSUE_DISPOSITION_MAP.md](TheTell/files/ISSUE_DISPOSITION_MAP.md) — all 33 issues from the 2026-07-03 claude.ai audit mapped to their handling (calls/PRs/pregrades) + the 4-nod rider. Beside it: the exported `fingerprint_findings.csv` + `fingerprint_review_summary.md` (Blocker #0 resolved 2026-07-07).
- _Knowledge-base wine guides_ in `Blind-Tasting-App/knowledge-base/` — grape-specific reference content.

## The Cave (`the-cave/`)

- [the-cave/CAVE_SYSTEM.md](the-cave/CAVE_SYSTEM.md) — **unified system vision/spec**: one data spine → 3 surfaces (print binder LIVE · staff console TO BUILD · kiosk tête de cuvée) + editorial layer. Console maintenance/capacity/optimization-brain spec, curation rules, build order. Read before any cave-system work.
- [the-cave/SETUP.md](the-cave/SETUP.md) — Xcode kiosk app setup, landscape lock, status bar.
- [the-cave/KIOSK_LOCKDOWN.md](the-cave/KIOSK_LOCKDOWN.md) — locking the kiosk iPad to the app: Guided Access now, Single App Mode from Apple Configurator, Apple Business enrollment (2026-11-12; unconfirmed items are marked).
- [the-cave/BOTTLE_PHOTOGRAPHY.md](the-cave/BOTTLE_PHOTOGRAPHY.md) — lightbox shoot spec, transparent PNG export.
- [docs/hardware/ipad-census.md](docs/hardware/ipad-census.md) — store iPad fleet: which gen does what (register / kiosk / spare), trickledown rule, Kyle's census to fill.
- [the-cave/PRODUCT.md](the-cave/PRODUCT.md) — kiosk product spec (2026-11-06 rebuild): etiquette direction, carousel browse → etiquette expansion, reads `cave_kiosk` only.
- [the-cave/design/directions.html](the-cave/design/directions.html) — the three direction comps Kyle chose from (etiquette ruled 2026-11-06).
- _Editorial layer: `the-cave/editorial/` — `verify.mjs` (deterministic gate) + `land.mjs` (writes `cave_editorial`); worker briefs in `.ferry/2026-11-06-cave-editorial-*`. Migration package: `db-migrations/2026-11-06-0001-cave-editorial-kiosk/`._
- _Print menu generator: `the-cave/menu/generate-menu.cjs` (LIVE, see CHEATSHEET). v3 HTML = classification source, don't delete._

## Wine Atlas

- [thief-wine-atlas/ATLAS.md](thief-wine-atlas/ATLAS.md) — full project doc. WA AVAs + Loire 6-bucket taxonomy, exact INAO geometry, PostGIS-backed API. Current as of 2026-07-11 recovery.
- [thief-wine-atlas/BLUEPRINT.md](thief-wine-atlas/BLUEPRINT.md) — the build-out scoping: canon linkage (55/62 auto-match), grape backfill, stories pipeline, PMTiles, region expansion, integrations. Kyle's 5 open calls at the bottom.
- [thief-wine-atlas/recovery/2026-07-11/RECOVERY_NOTES.md](thief-wine-atlas/recovery/2026-07-11/RECOVERY_NOTES.md) — the un-parking: Session-4 diagnosis, fixes, Kyle's 2am exactness veto, Loire v2, open gates + rollbacks.
- [thief-wine-atlas/README.md](thief-wine-atlas/README.md) — Next.js bootstrap.

## Enrichment / scraping (`thief-wd-rescue/`)

- [inventory-intel/BUYERS_PREP.md](inventory-intel/BUYERS_PREP.md) — **NEW 2026-09-25. The buyer's prep v1 = the CASE-BREAK CHECK (Horizon B15b):** Kyle's ruling verbatim ("is the 3 case price 20% less than the last purchase… I buy three cases but it takes us 6 months"), what `scripts/buyers-prep-sheet.mjs` reads (Vehrs/Winebow `discounts` jsonb → `distributor_links` → active listings; last paid from GG receipts; paid-only pace, comped badge), the formulas (eligible best-of, ladder, inclusive months of supply, 20% + 3-month thresholds), the ceilings (SGWS deal price unreadable; Winebow has no multi-case tiers; wines we carry only), controls (Caparzo 13190), how to run, v2 = the Reorder tab column decided by the `already knew` CSV. Sheet lands OUTSIDE git in `~/Documents/buyers-prep/`.
- [inventory-intel/STOCK_HEALTH.md](inventory-intel/STOCK_HEALTH.md) — **NEW 2026-07-25. The stock-health triage (B15 Phase 1 live artifact)**: velocity × on-hand classes (actionable dead $84.7k / kill list $136k / overstocked $203k), the three classification laws that survived spot-check (vintage-family, pour-fed, billboard), re-run command, and the dynamic-surface proposal (reorder-tab color chips + `m_movers` ratify). Generator: `thief-wd-rescue/stock-health-report.mjs`.
- `inventory-intel/BUNDLE_TWIN_AUDIT_2026-09-05.md` — sibling audit of every `shopify_products` reader vs the spirits bundle parent/component twin (66 objects; 7 fixed, 4 need a migration, 2 need Kyle's EAN ruling)
- [thief-wd-rescue/FTI_PHASE2_RECONCILE.md](thief-wd-rescue/FTI_PHASE2_RECONCILE.md) — **NEW 2026-07-25. FTI phase-2 apply state of record**: reconcile of the two independent 07-24 dry-runs (zero conflicts; agreed core 908 orders/$55.5k), apply 303/908 landed + morning runbook to finish, residual Kyle rulings (A-only/B-only classes, 14 merge candidates). Method detail: `thief-wd-rescue/FTI_PHASE2_DRYRUN.md`; diff tool: `fti-phase2-reconcile.py`.
- [thief-wd-rescue/DOMESTIC_FLESH_BRIEF.md](thief-wd-rescue/DOMESTIC_FLESH_BRIEF.md) — **NEW 2026-07-13. Dispatch brief for the domestic-winery FLESH pass** (Opus lane): design calls made (bios ride `enrich-local-wineries.mjs`; wine flesh → `domestic_winery_catalog` `source='winery_page'`; JS sites skip-and-list), laws, allocation-house priority targets, definition of done. Spine state: `winery_sites` + log.md 2026-07-13 entry.
- [thief-wd-rescue/IMPORTER_CAMPAIGN.md](thief-wd-rescue/IMPORTER_CAMPAIGN.md) — **NEW 2026-07-10. The importer-warehouse campaign state board**: ranked queue (repository-first — overlap sequences, never trims), per-importer status, Kyle gates, crawl progress, scoop worklist. Any terminal resuming the lane reads this first. Match-specs for fresh bridges: `thief-wd-rescue/specs/`.
- [thief-wd-rescue/CANON_SHOPIFY_BACKFILL_BRIEF.md](thief-wd-rescue/CANON_SHOPIFY_BACKFILL_BRIEF.md) — **NEW 2026-07-09, the canon→Shopify geography backfill spec of record**: target state per wine (tags full-tree / `shopify.*` taxonomy-law / `custom.appellation`), Burgundy 4-axis + dual-role appellation rules, GID census results, apply protocol + gates. Dry-run CSVs + ratify surface: `_enrichment-spike/geo-backfill/`. Generator: `thief-wd-rescue/scripts/canon-geo-backfill.mjs` (dry-only until ratified).
- [thief-wd-rescue/ENRICHMENT_PIPELINE.md](thief-wd-rescue/ENRICHMENT_PIPELINE.md) — all enrichment scripts, run order, ScraperAPI gotchas.
- [thief-wd-rescue/scripts/JUDGE_PROMPT.md](thief-wd-rescue/scripts/JUDGE_PROMPT.md) — **NEW 2026-09-04.** The ONE instruction file every corpus-judge agent reads (control and fleet alike): what `match` means, the identity-bearing vs ignorable word classes, `match|unsure|absent`, exact output shape. Judge tier is SONNET low — Haiku failed the blind control (`corpus-match-llm.mjs --control` / `--grade-control`).
- [thief-wd-rescue/_corpus-out/RUN-2026-09-04-corpus-linkage.md](thief-wd-rescue/_corpus-out/RUN-2026-09-04-corpus-linkage.md) — **2026-09-04 captain-run report**: control scores (Sonnet PASS / Haiku FAIL), 156 judged → `llm-proposals.csv`, 44 verified cuvée candidates, the 26 dangling links, price actual vs forecast, Kyle's next moves in order.
- [thief-wd-rescue/GAP_DECK_DESIGN.md](thief-wd-rescue/GAP_DECK_DESIGN.md) — **Gap Deck design of record (2026-09-05; §8 RATIFIED, steps 1–4 BUILT, migration held at Kyle's gate)**: Kyle closes the corpus↔Shopify gap one card at a time — ratify (Enter) / pick (1–5) / scoop (Grab Wine) — with the `capture_catalog` feeder silo so scoops survive the nightly polaroid rebuild. ⚠ **Read §10 first** — the build overturned five things in §4/§6/§7. Ferry: `.ferry/2026-09-05-gap-deck.md`.
- [thief-wd-rescue/ALLOCATION_ARCHITECTURE.md](thief-wd-rescue/ALLOCATION_ARCHITECTURE.md) — allocation offer workflow + scoring formula.
- [thief-wd-rescue/IMPORTER_PRODUCERS_SUMMARY.md](thief-wd-rescue/IMPORTER_PRODUCERS_SUMMARY.md) — 5,920 producer profiles across 48 importers.
- [thief-wd-rescue/docs/shopify-taxonomy-values.md](thief-wd-rescue/docs/shopify-taxonomy-values.md) — Shopify metaobject values reference.
- [thief-wd-rescue/BULK_OPERATIONS_SCOPE.md](thief-wd-rescue/BULK_OPERATIONS_SCOPE.md) — Shopify Bulk Operations API scope + spike results. **SHIPPED 2026-07-11:** full product sync migrated to bulk query (no `first:N` caps, ~2.5 min); delta stays paginated; bulk *mutation* helper still opportunistic (UPC backfill / image repair are its first customers).

## Marketing (`marketing/` — promoted out of `WebsiteModification/` 2026-07-14)
- OND Lane 1 drafts (2026-10-11, nothing sent): `marketing/thanksgiving/2026-package-DRAFT.md` (8-bottle pairing guide, all verified in WW stock; 11-12 email plan; hours line "Kyle confirms") · `marketing/holidays/2026-nov-dec-slate-DRAFT.md` (18 tier A/B rows: IG caption + Vestaboard board each, boards 6×22-checked) · `marketing/briefing/issue-08-DRAFT.md` (Fri 10-16; 5 Kyle-checks on top; standing "Get on the list" texts section added).

- [marketing/README.md](marketing/README.md) — **NEW 2026-07-14. The cupboard map — start here for any marketing work.** What lives where (brand · letters · campaigns · print · signage), the four house rules, and the open Klaviyo self-serve thread. Marketing used to live inside the Shopify *theme* repo, which is why nobody could find it.
- `marketing/business-cards/README.md` — staff business card template (house card, deco gold back, `build.sh` → print PDFs w/ bleed); first printed card = Dan Hoffheins 2026-10-29
- `marketing/newsletter/README.md` — the printed monthly (Kermit Lynch shape, Thief voice): spec, build command, content rules, open decisions; issues in `marketing/newsletter/YYYY-MM/` (`issue.html` + `issue.pdf`). Issue one = 2026-10 MOCKUP. `2026-11/beaujolais-pack.md` = the November Beaujolais 6-pack (bottles, cost, margin, tasting-overlap rules).
- [marketing/BRAND.md](marketing/BRAND.md) — **NEW 2026-07-14. Read before putting the logo on ANYTHING.** The naming law (**a lockup is named for the background it sits ON, never for the colour of its own letters** — the old `thief-logo-white-with-gradient-clear-background.png` has *black* letters, which is precisely why it kept getting grabbed wrong). The canonical set now lives in `marketing/logos/` (`-light` / `-dark` / `-on-black`). Carries two traps: the loose root PNG is **load-bearing for 44 tasting-menu files — never "tidy" it**, and email needs a *hosted* URL, not a file path.
- [marketing/shopify-notifications/PASTE_PACK.md](marketing/shopify-notifications/PASTE_PACK.md) — **NEW 2026-08-15.** Shopify customer-notification audit: all 7 templates are untouched stock defaults, **no API surface exists** (re-proven live; the Shopify in-admin AI hallucinates one), read path = "Send test email" → kyle@ → M365 MCP. Carries the 3-paste pack (brand Customize + weather-hold copy) and the un-built phase-2 per-order hold email.
- [marketing/klaviyo/upload-assets.py](marketing/klaviyo/upload-assets.py) — **NEW 2026-07-14.** Hosts the display-type artwork in Klaviyo's image library and writes `image-manifest.json`. **This is what the whole email program was blocked on** — the Briefing/Rickhouse templates had been finished since 07-11 with every `<img src>` reading `PASTE-KLAVIYO-URL-FOR-…`. Idempotent; `--dry` previews. Email cannot read a local path, and Boucherie can never be live text, so hosting artwork is load-bearing, not admin.
- [marketing/INSTAGRAM_VOICE.md](marketing/INSTAGRAM_VOICE.md) — **DRAFT 2026-08-30, Kyle ratifies. The Instagram register** — a different room from the menu voice: valley first, theme once, facts light (no recital, not no facts), invite-not-CTA close, two-layer hashtags (topic furniture + Kyle's running joke tags — **reuse, never invent**). Derived from 15 real captions in `marketing/instagram/voice-samples.md`, not from the menu voice.
- [marketing/instagram/SETUP.md](marketing/instagram/SETUP.md) — **The Instagram posting rail — design of record. RAIL BUILT 2026-09-04** (§3: `post.py` approve/`--publish`, `publish-due.sh`, the carousel + story-fit + caption-file + receipt rulings). **TOKEN LIVE 2026-09-03** (system user, never expires; env `IG_ACCESS_TOKEN`/`IG_USER_ID`/`FB_PAGE_ID`; the dead Instagram-Login/tester route is recorded in §1 so nobody walks it again), the verified can/can't table (feed + stories + reels yes; **stickers NO; location tag by API NO** — in-app yes since 09-03), the rulings on Facebook/Threads/X/WhatsApp/reels, cadence (§6a), phases. Scripts: `verify-token.sh` · `probe-location.sh` · `pull-archive.py`.
- [marketing/INSTAGRAM_VOICE.md](marketing/INSTAGRAM_VOICE.md) — **NEW 2026-09-03, RATIFIED.** The Instagram register, written from the 2022-26 archive (387 posts): the Thursday caption shape, measured mechanics (ellipsis not dash, 1-3 `!`, 0 emoji), the two-layer hashtag law (topic + Kyle's running-joke tags — reuse, never invent), the negative space (no tasting notes, no salesman words, no poster recital), 14 verbatim reference captions, the engagement read (shop-life posts out-perform tastings; carousels 1.5×), and the tic ledger the rail enforces.
- [marketing/instagram/post.py](marketing/instagram/post.py) — **NEW 2026-09-04.** The Instagram rail. Approve (reads `thursday_tastings`, builds the 9:16 story fit, hosts on the Klaviyo CDN, creates the Graph container, lints the caption, writes `queue/<date>-<kind>.json`) → `--publish` (asks, publishes, reads the permalink back, PATCHes the receipt). `--restamp` fixes a failed receipt without re-posting. Never writes a caption, never tags a location, never auto-fires.
- [scripts/analytics/second-visit/README.md](scripts/analytics/second-visit/README.md) — **The Second Visit (2026-09-26, read-only build):** the first 120 days decide who becomes a regular — WD-era on-ramp curve (second visit ≤120d → 33–42% reach 10 visit-days; >365d → 8%), first-basket bands (flights-first 0.5%), and the live list of 598 one-visit Shopify customers still in window (WD-era + pseudo excluded). `queries.sql` + `build.mjs` (controls a/b/c throw on failure) → `out/second-visit.html` (gitignored: names) → https://claude.ai/artifact/EY9KpxwY9K394KPUQMkNfD. Brief: `.ferry/2026-09-26-second-visit-window.md`.
- [scripts/analytics/palate-map/README.md](scripts/analytics/palate-map/README.md) — **The Palate Map (2026-09-26, read-only build):** producer co-purchase constellation from 8 years of receipts (lift ≥5, n ≥20, burst-filtered), d3-force page with a habit layer (buyers · repeat share · in stock now). Brief: `.ferry/2026-09-26-palate-map.md`.
- [.ferry/2026-11-04-zanzibar-quant-bench-brief.md](.ferry/2026-11-04-zanzibar-quant-bench-brief.md) — **Quant bench, the 128 vs 256 Mac decision (2026-11-04, DONE, $20.95 RunPod):** Qwen 4-bit (128GB box) loses visibly vs hosted; Qwen 8-bit + DeepSeek 2-bit (256GB box) each one narrow miss from hosted. Verdict: 256 is the box; 128 is a worse tier, not a cheaper one. Page: `zanzibar/bench2/bigger-box-bench.html`.
- [.ferry/2026-09-21-champagne-2c-shapes-and-release.md](.ferry/2026-09-21-champagne-2c-shapes-and-release.md) — **Champagne drill-down: state + Kyle's release card + the shapes worker's recipe (2026-09-21).** 2a villages CSV and 2b producer→village built and verified; 2c tree/placer staged (Rebuild lane, Kyle's `--apply`); shapes bake = one `opus48-worker` day. Sitting sheet: `ThickAsThieves/passport/out/champagne-gavel.html`.
- [.ferry/2026-09-04-instagram-rail-REVIEW.md](.ferry/2026-09-04-instagram-rail-REVIEW.md) — **review ask for Fable**, written by the maker: scope, migration sha, the 8 bugs already found and fixed (so rounds aren't spent re-finding them), and the 5 judgment calls I most want attacked. Maker-never-grades-itself; none of its framing is settled.
- [.ferry/2026-09-04-statusline-findings.md](.ferry/2026-09-04-statusline-findings.md) — 4 leads in **another lane's** uncommitted work (`harness/statusline.sh` ×3, duplicated `public.wd_products` in `db-docs/schema.json`), swept in by a whole-tree review. Not touched; handed over.
- [marketing/instagram/story-card.py](marketing/instagram/story-card.py) — **NEW 2026-09-04.** Cuts a 1080×1920 words-only story in house type (Roslindale; Block is full caps and can't set a sentence). Ground is a blurred photo on purpose — BRAND.md rules three colours and no flat dark ground. Reuses `tasting-graphic.py`'s `cut()` so plates stay outlined artwork, never a font file. Balanced wrap (no orphan last word) + vertical scrim.
- [marketing/instagram/publish-due.sh](marketing/instagram/publish-due.sh) — **NEW 2026-09-04.** Fires queued posts whose `publish_at` has arrived; 18h stale window; `--dry`. Only `queue/*.json` is live — `done/` + `rehearsal/` are inert. launchd plist STAGED, NOT LOADED (census row 38).
- [marketing/instagram/captions/](marketing/instagram/captions/) — one caption file per post, `<date>.md` / `<date>-story.md`. **Whole file is the caption verbatim**; an optional `---` block at the top is notes. `2026-09-03.md` is the worked example. Written by a terminal against `INSTAGRAM_VOICE.md`, vetoed by Kyle.
- [db-migrations/2026-09-04-0001-tasting-instagram-permalink/](db-migrations/2026-09-04-0001-tasting-instagram-permalink/) — **NEW 2026-09-04, NOT RELEASED.** One nullable column, `thursday_tastings.instagram_permalink` — the feed-post receipt, same shape as `blog_published_at`. Surveyed (Gate 0.5), rehearsed rung 1 with rollback post-checked, 7 in-file assertions incl. a positive control. Gate 5 (Codex) required by the DO block; Kyle releases.
- [db-migrations/2026-09-06-0001-shop-product-detail/](db-migrations/2026-09-06-0001-shop-product-detail/) — **NEW 2026-09-06, NOT RELEASED.** `get_shop_product_detail(handle)` — one bottle for the TaT drill-in sheet: every `get_shop_products` field at the same tier behind the same 21+ gate, plus `producer_bio/tasting_notes/winemaker_notes/importer/scores_json/abv/country/body_html` and the best-confidence corpus row. **Never `staff_notes`; no client-supplied tier.** Survey found the brief's `is_listing` filter isn't what the live list uses (mirrors the list instead) and that corpus `soil` reaches 3,853 chars. Three prod rung-1 rehearsals; **Gate 4 became empirical after Codex proved the `pg_policy` detector blind to PUBLIC (OID 0 has no `pg_roles` row)** — now `SET LOCAL ROLE authenticated` with a real non-staff claim: RPC 1 row, direct table read 0. Kyle releases.
- [.ferry/2026-09-21-invoice-lines-to-corpus.md](.ferry/2026-09-21-invoice-lines-to-corpus.md) — **EXECUTED 2026-09-21** (945 live GG invoice lines keyed to the corpus; step 3 Jev calibration also run — see `docs/rules/enrichment.md`). Kept as the record of the survey + write-scope law.
- [.ferry/2026-09-21-catalog-refresh-stale-portals.md](.ferry/2026-09-21-catalog-refresh-stale-portals.md) — **NEW 2026-09-21, fired same night.** Manual refresh of the four distributor catalogs untouched since 05-22 (Provi's 7 houses, Columbia, Odom, Grape): scripts exist, schedule never did; per-house ±30% row-count verifier; Grape + any MFA = Kyle-at-desk.
- [.ferry/2026-09-06-scan-to-bottle-report.md](.ferry/2026-09-06-scan-to-bottle-report.md) — **NEW 2026-09-06.** Morning report for the overnight scan-to-bottle lane: the 5 numbered steps needing Kyle (veto → release → Apple/Play logins → `vercel deploy` with a before/after curl table → scan a real shelf tag), what shipped on both clients, the pre-existing `deleted_at` list bug found and deliberately not fixed, the Codex reconciliation, and the two instrument-silence near-misses (a mislabelled capture; `simctl io screenshot` refusing to overwrite while printing success).
- [db-migrations/2026-09-04-0002-rail-tab-customer/](db-migrations/2026-09-04-0002-rail-tab-customer/) — **RELEASED 2026-09-04, read-back green.** `wine_bar_tabs.shopify_customer_id` + `customer_name` + `customer_linked_source` (+ two CHECKs): somewhere for a bar tab to hold WHO it is for, so the Rail's push-to-cart can `setCustomer()` at the register. Survey measured the problem first — the Rail attaches a customer 39.2% vs the same store's counter at 53.0% — and killed the obvious design: `guest_name` is a seat label (176 of 221 distinct values are one token), not identity. **Gate 5 (Codex) HELD it and was right — the money bug was `clearCart()` not removing the cart's customer, in the extension, not the SQL** (`LESSONS_LEARNED.md`). Rehearsed rung 1, rollback + its refusal path exercised, assertions proven able to fail. Kyle releases.
- [marketing/instagram/posts-archive.json](marketing/instagram/posts-archive.json) — every @thethiefshop post (1,160, 2018-2026) with caption/type/likes/comments/permalink; `.csv` sibling. Re-pull: `pull-archive.py`. Frozen snapshot of 2026-09-03.
- [marketing/MARKETING_STRATEGY.md](marketing/MARKETING_STRATEGY.md) — **RESHAPED 2026-07-03** into the marketing operating system: channel map (SMS=doorbell / email=detail / app=showroom / IG=top-of-funnel), Kyle's 3-5hr weekly rhythm, 30-day activation sprint, app-launch merge lane, paid-spend gate. Read first for any marketing work.
- [WebsiteModification/README.md](WebsiteModification/README.md) — **NEW 2026-07-25. Read before touching the Shopify theme.** Repo identity (`pottsy17/thief-website` — theme + copy live together, never a separate theme repo) and the trap that matters: `live-theme/` and `candidate-theme/` were pulled by **different tooling**, so `diff -rq` cries 29 changed files when only **3** differ in substance. Carries the three real diffs — incl. the candidate **missing the age-verification app block**, which is why the box-sizing rule ships as one `custom_css` line pushed to the LIVE theme, never by publishing the candidate. Also: `live-theme/config/settings_schema.json` is malformed (trailing comma, vendor's file, live and working).
- [WebsiteModification/thief-website-copy.md](WebsiteModification/thief-website-copy.md) — Shopify homepage + about + contact copy.
- [WebsiteModification/WEBSITE_AUDIT_2026-06-24.md](WebsiteModification/WEBSITE_AUDIT_2026-06-24.md) — **current** verified site audit (supersedes 2026-03-22). Spirits page shipped, homepage placeholder fixed; open: Tastings/Events nav dupe, Visit page needs both addresses.
- [WebsiteModification/WEBSITE_AUDIT_2026-03-22.md](WebsiteModification/WEBSITE_AUDIT_2026-03-22.md) — site audit, broken links, content gaps (superseded).
- [WebsiteModification/sparkling-tasting-2026-03-05.md](WebsiteModification/sparkling-tasting-2026-03-05.md) — sparkling tasting menu.
- [marketing/PREFERENCE_CENTER.md](marketing/PREFERENCE_CENTER.md) — **NEW 2026-07-11.** How people choose what we send: 4 topics (Briefing / Rickhouse / Thursday Tastings / Rare & Allocated), the Klaviyo build order, and the load-bearing rule — **The Rickhouse is opt-in and never blasted to the 6,054** (Briefing #1's anti-abuse strike is why). The preference page is the *launch mechanism* for the second letter, not a nice-to-have. Read before adding any new email list or send.
- [marketing/_brand/README.md](marketing/_brand/README.md) — **NEW 2026-07-11.** The wordmark factory: `wordmark.py` renders Boucherie as **outlined artwork** (SVG paths + PNG), so our Desktop-only font licence never needs the yearly Web/App embedding subscriptions. **Read before setting ANY Thief display type** — headline, masthead, shelf talker, sign, app. 29 marks generated; specs in `_brand/specs/`, output in `_brand/out/`. Also carries the **truncated-lockup post-mortem**: 5 of the 13 hand-authored SVGs in `site-glow-up/lockups/` had silently-cut path data ("JOIN THE LIST" → *JOIN*) — valid XML, no error, caught only by eye. Never hand-write path data.
- [scripts/anon-definer-tripwire.sql](scripts/anon-definer-tripwire.sql) — **NEW 2026-08-22.** Companion to `rls-audit.sql` for the surface its §7 re-triages noisily: anon-executable `SECURITY DEFINER` functions. Carries a **reviewed manifest of 26** (verdicts from `docs/anon-securitydefiner-replay-2026-08-22.md`) so it stays SILENT until a NEW or DROP+CREATE-re-armed anon-reachable definer function appears — clean = zero `ALARM` rows. Read-only pg-catalog query, `get_pour_menu` as the control. To add a function: replay it with the anon key, then append `(name, verdict)`. Run via the MCP/weekly hygiene like `rls-audit.sql`.
- [scripts/lint-truncation.py](scripts/lint-truncation.py) — **NEW 2026-07-12.** Sweeps every machine-data file (svg/csv/json/sql) for the **cut-but-still-parses** bug class. The check is never "does it parse" — that's what the broken lockups passed — it's "is the content all there": SVG ink-bbox must fill the declared viewBox · CSV rows must match the body's modal field count · JSON must parse · SQL must end terminated. HIGH/LOW tiered. **Run after any bulk CSV/SVG generation.** $0, no model.
- [harness/README.md](harness/README.md) — **NEW 2026-09-01, the tracked Claude Code harness layer.** `hooks/` (every guard + discipline hook), `commands/` (`/ask-codex` `/lint-wiki` `/seat`), `agents/`, `statusline.sh`, and `settings.hooks.json` (registry of record). Reverse-symlink pattern: real files here, `~/.claude/...` links into them via `install-mac.sh`; the PC registers the node guards directly. Portability rules (repo root from own path, `HOME||USERPROFILE`, `os.tmpdir()`), the fixture-gated edit rule for the SQL guard, and the **seats** rail (human lane names in live-lanes, the footer, and the tab title). Closed MACHINES.md gaps #1 (repo side) + #2.
- [scripts/pc/README.md](scripts/pc/README.md) — **NEW 2026-08-31, the Windows doppelganger rail.** PowerShell ports of the wiki-discipline hooks (auto-commit-with-push + log-nudge) + idempotent `setup-doppelganger.ps1` for the home PC. Since 2026-09-01 also registers the three node guards from `harness/hooks/` + carries `seat.ps1`; the remaining does-not-port list is the bash hooks, hygiene-lint, board hooks, sounds. Mac hooks stay authoritative; secret-pattern lists synced by hand.
- [marketing/_brand/compact-mark/README.md](marketing/_brand/compact-mark/README.md) — **NEW 2026-08-08.** The compact-mark sitting (BRAND.md §7's answer-in-waiting): 3 directions (Boucherie-T lozenge / sunburst crest / keyhole) + references + pre-filled call, awaiting Kyle's grade of `contact-sheet.png`. Masters are `currentColor` single-ink; rebuild via `build_directions.py`.
- [marketing/stamps/README.md](marketing/stamps/README.md) — **NEW 2026-08-07.** Rubber-stamp die art + the house stamp canon: vendor (stampmaker.com `WOOD_50X3.75` $56.90), the Spirits & Beer bag stamp (`build_spirits_stamp.py` — badge + big "WALLA WALLA, WA" truncated from the designer's own WASHINGTON glyphs), and the trap: **stampmaker picks the stamp product from CROPPED ink size, so die art ships with zero margin.**
- [marketing/_brand/TYPE_SYSTEM.md](marketing/_brand/TYPE_SYSTEM.md) — **NEW 2026-07-11.** Which of the 16 Boucherie cuts to use, and why. The family sorts **mechanically, by what each cut does with lowercase**: Block = full caps (*name* face) · Sans = small caps (*label* face) · Flared = true lowercase (*reading* face) · Cursive = script (*voice* face). Two load-bearing calls: **Block's power is scarcity — Sans should carry the repeating furniture**, and **Flared is a text face we'd been asking to shout**. Read before choosing a face for anything.
- _Campaign folders:_
  - [briefing/](marketing/briefing/) + [rickhouse/](marketing/rickhouse/) — **The two letters.** Briefing = weekly wine (the default list); Rickhouse = monthly booze (opt-in only, starts at zero). Each file IS its living template — copy the newest issue, refill the SLOT blocks. Both gated on the Klaviyo image upload + preference-page URL.
  - [cayuse-weekend-2026/](marketing/cayuse-weekend-2026/) — completed April campaign collateral.
  - [spring-release-2026/](marketing/spring-release-2026/) — May 1-3 campaign. Partial execution; no photographer.
  - [a-frame-sign/BRIEF.md](marketing/a-frame-sign/BRIEF.md) — sidewalk sign design brief.
  - [window-decals-2026/BRIEF.md](marketing/window-decals-2026/BRIEF.md) — **NEW 2026-07-03.** Both storefronts, specs locked (gold/black/red-21+ Oracal 030); blocked on Kyle's measurements + hours + .ai files.
  - [briefing/](marketing/briefing/) — **NEW 2026-07-03.** "The Thief Briefing" weekly Klaviyo email: issue-01 draft (md) + email-safe HTML (the file IS the living template — copy newest issue, refill SLOT blocks).
  - [sms-invite/](marketing/sms-invite/) — **NEW 2026-07-03.** SMS opt-in program collateral: invite email HTML, register QR card (print-ready), QR assets, NOTES.md. `/pages/texts` LIVE; gated on Klaviyo form URL-targeting.
  - [app-launch/](marketing/app-launch/) — **NEW 2026-07-03.** DORMANT launch package (email + SMS doorbell copy + IG reel script + fire-day checklist). Fire when Apple approves TAT.
  - [site-glow-up/BRIEF.md](marketing/site-glow-up/BRIEF.md) — **NEW 2026-07-03.** Verve-gap teardown (same theme! gap = art direction) + Boucherie lockups v1 BUILT (`lockups/`) + adopted art-direction spec (`art-direction.md`: shot list, grade, Firefly prompts).
  - [blade-sign/BRIEF.md](marketing/blade-sign/BRIEF.md) — **NEW 2026-07-03, PARKED.** Corner theatre blade sign: WW code caps at 8ft/32sqft external-lit (verified); ~$8-18k; first action = Development Services call (questions in brief).

## Plans / specs (`docs/superpowers/` — dir name is historical; the plugin is gone, the artifacts stay)

- [plans/2026-03-25-thick-as-thieves-phase-1.md](docs/superpowers/plans/2026-03-25-thick-as-thieves-phase-1.md) — TAT Phase 1 implementation plan.
- [plans/2026-03-26-thick-as-thieves-checkout-errors-brix-webhook.md](docs/superpowers/plans/2026-03-26-thick-as-thieves-checkout-errors-brix-webhook.md) — TAT checkout + Brix webhook plan.
- [plans/2026-06-23-tat-drops-new-vintage-rail.md](docs/superpowers/plans/2026-06-23-tat-drops-new-vintage-rail.md) — **NEW 2026-06-23.** TAT new-vintage "drops" rail v1 build spine (backend RPCs, DOB age-gate, push fan-out) + scope/decisions. Built + Codex-red-teamed; full arc in `log.md` 2026-06-23.
- [plans/2026-03-29-vestaboard-automation.md](docs/superpowers/plans/2026-03-29-vestaboard-automation.md) — vestaboard automation plan (original build).
- [plans/2026-07-17-vestaboard-rework.md](docs/superpowers/plans/2026-07-17-vestaboard-rework.md) — **Vestaboard rework brief (Opus execution, half-day):** cron → pg_cron (GH Actions starves the schedule), parity alternation, layout engine v2 (auto/waterfall/verbatim + composer fix), content-pool expansion (Now Spinning, WOTD, beer, arrivals, staff pick). Ratified by Kyle 2026-07-17/18.
- [plans/2026-07-18-shopify-door-sales-backfill.md](docs/superpowers/plans/2026-07-18-shopify-door-sales-backfill.md) — **Shopify door-sales backfill spec (Opus execution, money path — full review law):** ~227 pre-2026 tasting/door days (~$221k) → Xero, Kyle's proven 3-line invoice model, Lane E rail pattern reused. Scoped strictly < 2026-01-01.
- [AUTOMATION_ROADMAP.md](thief-ops/AUTOMATION_ROADMAP.md) — **NEW 2026-07-26 (ultracode survey).** The what's-next layer over cron-census: ranked tiers from "Kyle kicks it off" → scheduled. Tier 1 = SGWS rail on zanzibar (nightly scrape + closeout/new-arrival/DP diff); Tier 2 = six-report pack from existing scripts; Tier 0 = monitor debt (5 jobs still logless; drops-janitor/claim-sweep registry hole FIXED same session). 12-ruling gavel list at the bottom; awaiting Kyle.
- [BRIEF_TASTING_BLOG_WATCHDOG.md](thief-ops/BRIEF_TASTING_BLOG_WATCHDOG.md) — **NEW 2026-07-18. Opus execution, ~1hr.** Blog-drift watchdog: `blog_published_at` receipt column + stamp in `tastings/create-blog-post.mjs` + weekly pg_cron→edge-fn nudge email on the existing digest/Resend rail. Never publishes (hero graphics are Kyle-local); registered in `v_cron_health`. Greenlit by Kyle 2026-07-18.
- [BRIEF_STAFF_UI.md](menu-tool/BRIEF_STAFF_UI.md) — **BUILT 2026-07-28** (written 07-25 as a scope-ruling brief; now the build log of record, newest round at the bottom). The staff web UI: **the browser replaces the Excel sheet** — `npm run menu -- --ui` reads today's lineup off live POS, staff swap/add/take off, picked rows carry a Shopify product id so the matcher never fires. Also: localhost-only + OS-account boundary · never-halt extended to a closed tab · one-run-at-a-time enforced by the port · Pacific date pinned. Read the **Defect found by Kyle in first use** section before writing any browser UI here — *curl proves the endpoint, not the button.* Open: no real `--ui` menu run yet, parity verifier unrun, out-of-stock ruling pending.
- [shop-query/README.md](shop-query/README.md) — **NEW 2026-08-14.** The Fawkes Customer Query tab as a staff web app (`~/code/shop-query/`, private repo `pottsy17/shop-query`): loopback `:3030` on the shop box, `claude -p` on the second (Pro) account, **no `ANTHROPIC_API_KEY`**. Deterministic SQL filter + code ranking → top 120 candidates → one model call. Extracted from the now-archived `golden-goose/shop-tool/`; Menu stayed with `menu-tool`, Ask deliberately dropped.
- [shop-query/STAFF_BOX_SETUP.md](shop-query/STAFF_BOX_SETUP.md) — **NEW 2026-08-14, box runbook — ⏳ PENDING Kyle at the box (~25 min).** Mirrors the menu-tool rail exactly: second deploy key (GitHub keys are globally unique — menu-tool's cannot be reused), clone as `menusvc`, two LaunchDaemons, Phase C escalation controls. Gate before installing the daemon = `npm run smoke` (proves the effective model + that tools are off).
- [STAFF_BOX_SETUP.md](menu-tool/STAFF_BOX_SETUP.md) — **NEW 2026-08-03, the staff-box deploy runbook** (amends REQUIREMENTS §Deployment): box's `claude` signs into the second (Pro) account — never Kyle's Max; OS account split `menusvc`/`thief-staff` is the wall; tool verified wiki-free (both spawns `cwd: tmpdir()`); Phase C escalation tests are the controls. Execution started same day.
- [REFACTOR_NOTES_2026-11.md](menu-tool/REFACTOR_NOTES_2026-11.md) — **NEW 2026-11-01, notes not a plan.** What menu-tool should change so pour analytics stop reconstructing history: append-only edition history (pour_menu overwrites today), stuck-list-first picker, one slot for new arrivals, tie burn box to the run, backfill bottle tags on the 61 pre-May pour products. Source: the Glass Report build.
- [BRIEF_WHATS_POURING.md](menu-tool/BRIEF_WHATS_POURING.md) — **NEW 2026-07-18. Opus execution, ~half-day.** Machine + glass lineup → Shopify metaobjects (written by menu-tool after the POS confirm) → "What's Pouring" theme page on thiefshop.com; TAT reads the same rows via Storefront API later (PUBLIC_READ set day one). Pour products stay POS-only. Till-adjacent → both-vendor review.
- [plans/2026-06-06-menu-tool-build.md](docs/superpowers/plans/2026-06-06-menu-tool-build.md) — **Menu tool (Spec A) build plan, ready to execute in a fresh window.** 21 TDD tasks: scaffold `~/code/menu-tool/` → shared core (steal lookup-v2/pricing/shopify) → match+price → layout-fit (headless-Chrome, title subtract + note-trim) → cold-facts+voice notes → orchestrator+checkpoint → migrate+replay Edition 116. Implements `specs/2026-06-06-menu-tool-redesign-design.md`.
- [plans/2026-06-11-command-deck-mission-control.md](docs/superpowers/plans/2026-06-11-command-deck-mission-control.md) — Command Deck mission-control build plan (13 TDD tasks, CLI-proves-before-pixels). **Executed + shipped 2026-06-12.**
- [specs/2026-06-11-command-deck-mission-control-design.md](docs/superpowers/specs/2026-06-11-command-deck-mission-control-design.md) — Command Deck Mission Control + re-entry briefs design (state-machine semantics, tiered alerting, deterministic briefs) + the staged strategy (Mac-for-Kyle → Operator's Stack essay test → demand-gated Pro). **Shipped 2026-06-12**; holds the JSONL format verification notes.
- [specs/2026-06-12-command-deck-dress-up-interaction-design.md](docs/superpowers/specs/2026-06-12-command-deck-dress-up-interaction-design.md) — Command Deck dress-up + interaction design (theme tokens/presets/fonts/glass · sticky rename · terminal-agnostic click-to-surface) + the resolved Ghostty `ai-title` spike. **Shipped 2026-06-12/13.**
- [plans/2026-06-12-command-deck-dress-up-interaction.md](docs/superpowers/plans/2026-06-12-command-deck-dress-up-interaction.md) — Command Deck dress-up build plan (11 tasks, inline-executed not subagent-per-task). **Executed + shipped 2026-06-12/13.**
- [specs/2026-08-03-tat-android-v1.md](docs/superpowers/specs/2026-08-03-tat-android-v1.md) — **NEW 2026-08-03. TAT-Android v1 design spec — the artifact the August RN lanes execute from.** Parity target = TAT-iOS as shipped (9,612 lines / 5 tabs / 10 feature areas, inventoried). Rules the monorepo (`~/code/thief-mobile/`, npm workspaces, `apps/tat` + `packages/ui`), app id `com.thiefshop.tat`, Google+email auth (no SIWA on Android), FCM push, build order, honest price. ⚠ Carries the `send-push` platform-filter landmine — deploy that filter BEFORE any Android device registers.
- [specs/2026-03-25-thick-as-thieves-design.md](docs/superpowers/specs/2026-03-25-thick-as-thieves-design.md) — TAT design spec, brand identity, colors, fonts.
- [specs/2026-03-26-thick-as-thieves-checkout-errors-brix-webhook.md](docs/superpowers/specs/2026-03-26-thick-as-thieves-checkout-errors-brix-webhook.md) — TAT checkout spec.
- [specs/2026-04-07-enrichment-page-design.md](docs/superpowers/specs/2026-04-07-enrichment-page-design.md) — GG `/enrich` page design.
- [specs/2026-06-06-menu-tool-redesign-design.md](docs/superpowers/specs/2026-06-06-menu-tool-redesign-design.md) — **Menu builder redesign (Spec A), awaiting red-pen → implementation.** Stateless (Shopify = live DB, per-bottle cache extends `menu_notes`), cold-facts-locality-tiered notes, subtraction fit engine (rendered-width in headless Chrome, not char count), Approach A (deterministic engine, Claude severable / $0 without it). Supersedes `golden-goose/MENU_REFACTOR_PLAN.md`.
- [menu-tool/SPEC.md](menu-tool/SPEC.md) — **NEW 2026-06-18. THE canonical menu-tool source of truth** (consolidates the lossy lineage MENU_REFACTOR_PLAN→design→REQUIREMENTS→MATCHER_SPEC; wins on conflict). Locks: `menu_notes` = NOTE-ONLY (titles/sections/region DERIVE, overrides rare); titles = chiselled `buildWineTitle` ✅wired; sections = `categorize.ts` two-sided taxonomy ✅ported; region = `Appellation, Region/State` normalizer ✅built; notes waterfall (in-house → web LAST → **stage, human promotes, never auto-write Supabase**); north star = Claude-Code-free. Open: cold run + the capture/self-heal build.
- [menu-tool/REQUIREMENTS.md](menu-tool/REQUIREMENTS.md) — **(2026-06-18: consolidated into SPEC.md above; keep for the #1–10 requirement history.)** THE authoritative menu-tool intent/state doc (post-rebuild correction — success = what's DIFFERENT from the old two-panel thing, not "tasks pass").** Tracks all 10 requirements: #1–3 + #9 + #10-core **DONE 2026-06-08** (unify panels, single-source layout tokens, real autofit, `-p` notes off-API, in-house facts + cache-once); #10 distributor/web fetch tiers + Kyle's live test pending. Now the live record — supersedes the plan/spec above for current state.
- [menu-tool/MATCHER_SPEC.md](menu-tool/MATCHER_SPEC.md) — **NEW 2026-06-18. The locked matcher/lookup build spec** (a fresh session builds from it cold). Recall-first, vendor-anchored, vintage-weighted matcher — **CORE BUILT + WIRED, 15/16-raw proven (#9 alias-covered)**; the 16-input regression set; the titles (`buildWineTitle`) / sections (`categorize.ts` port) / `glass_pour_menu`-table decisions; and **PHASE 2 web note-sourcing ⚠️NOT BUILT** (region-aware: WA/CA→producer site, EU→US importer→`wine_producers`). Read alongside REQUIREMENTS.md's 2026-06-18 status banner ("✅DONE describes built, not works").
- [specs/2026-06-06-spec-c-staff-deployment-notes.md](docs/superpowers/specs/2026-06-06-spec-c-staff-deployment-notes.md) — **Spec C stub.** Staff tool = locked Claude Code terminal (harness-enforced permissions, managed settings + hook + non-admin kiosk), NOT a web wrapper. Verified lockdown recipe captured; full brainstorm later.
- [specs/2026-06-15-fire-compliance-ops-tool-design.md](docs/superpowers/specs/2026-06-15-fire-compliance-ops-tool-design.md) — **NEW 2026-06-15. DESIGN — GATED (no build until the business is a go post-Tim).** v1 MVP for the fire/life-safety compliance ops tool (companion to `docs/ideas/FIRE_COMPLIANCE_FEASIBILITY.md`). Internal ops system, NOT a BuildingReports SaaS competitor (WW has no filing rail to lose to). Next.js + Supabase + Vercel, $0 except optional SMS. Holds the data model (account→device→service-record→loaner→swap, with extinguisher class/rating/size load-bearing for swap compliance), the two-mode scheduler (due-engine two-clocks → extinguisher swap-queue + on-site appointment book w/ per-account hour constraints), scan/label flow (QR asset-labels via existing Zebra, online-only v1), 3 outputs, build-vs-buy + the one metered line (SMS). Scoped by two truths: zero accounts day one + WW is a 7-min town (efficiency = bench-batching via loaner swap, not routing).

## Pending session prompts (`docs/pending-prompts/`)

- [shopify-mirror-and-splice-prompt.md](docs/pending-prompts/shopify-mirror-and-splice-prompt.md) — store analytics Phase A (Shopify orders mirror). Ready to start.
- [wd-tender-import-prompt.md](docs/pending-prompts/wd-tender-import-prompt.md) — ✅ DONE 2026-06-08 (superseded). 8-year WD tender CSV loaded as the frozen `wd_raw_*` archive + deduped `wd_sales_orders` instead of the narrow plan. History only.

## TAT-pending (`docs/pending-tat-features/`)

- [occasion-planner/README.md](docs/pending-tat-features/occasion-planner/README.md) — Occasion Builder feature, ported from Fawkes. Adaptation notes inside.

## Staff SOPs (`docs/sops/`)

- [README.md](docs/sops/README.md) — the SOP program: six cards + weekly sheet + Monday review loop, ownership + cadence (SOPs die as orphaned documents). Card owners recorded here after the 2026-07-05 staff meeting.
- [cards.html](docs/sops/cards.html) — printable source: Opening / Closing / Quiet Time / Delivery (Receiving QA) / Dead Scan / **Mailing Address** (added 2026-08-21 — new card, print + laminate) + the weekly sheet. Edit here, reprint, relaminate.
- [sunday-meeting.md](docs/sops/sunday-meeting.md) — rollout plan + Kyle's talking notes for the 2026-07-05 staff meeting (red-pen round, owner assignment, the Tuesday promise).
- [manual/INBOX.md](docs/sops/manual/INBOX.md) — **NEW 2026-10-06.** Ops-manual chapters ("how you do the job", not checklists): Kyle dumps raw by voice, saved verbatim here, shaped into `manual/<section>.md` (5 questions). Crew red-pens, never authors.
- [manual/beer.md](docs/sops/manual/beer.md) — first chapter: cold-box facings + rotation (old cans into four-packs). Draft, open questions at the bottom.
- [manual/checking-in.md](docs/sops/manual/checking-in.md) — check-in rule for ALL product: the stacks are the reference, find each case on the invoice (not invoice→stacks); the Delivery card spot-check comes after. Wine layers owed.

## HR / employment (`docs/hr/`)

- [README.md](docs/hr/README.md) — the legal/HR layer: what each doc is, the three-layer staff-doc architecture (handbook → operations manual → cards), steal-downward-never-contradict-upward rule, pending GM-role counsel touch-up.
- [Thief-Employee-Handbook-May-2026.md](docs/hr/Thief-Employee-Handbook-May-2026.md) — **derived extraction; the `.docx` beside it is the counsel-reviewed version of record.** WSLCB/licensing rules, tasting-on-shift policy, conduct, pay/leave, confidentiality. Handbook §2.10 delegates opening/closing procedure to `docs/sops/`.
- NDA + Acknowledgment (`Thief-NDA-May-2026.*`, `Thief-Handbook-Acknowledgment-May-2026.*`) — signed at hire; `.md` copies are derived extractions.

- [LEAK_LANE_DESIGN.md](TheTell/benchmark/LEAK_LANE_DESIGN.md) — **DESIGNED 2026-08-15, deliberately NOT BUILT.** The Tier-1 answer-leak lane: root cause is `generate-suites.mjs:126` ending every coaching prompt with "The correct answer is X" while `:140` forbids naming it — a live prohibition that lost 48 of 61. Carries the 6-shape leak taxonomy w/ counts, the two-half fix (C2-strong + output filter), and **5 blockers from two independent adversaries** incl. a measured 29.5-33% over-block against our own 1,232 rows of certified tasting prose. Also documents that the suite metric scores leak-absence not quality (4 of the 13 "clean" rows are product failures). Read §5 before building. One Kyle ruling open.

## Gavel sheets (`docs/gavel-sheets/`)

_The standing gavel sheet: when the bench passes ~25 Your-Move cards or a window is closing, a terminal regenerates ONE bucketed decision page with my call pre-filled on every line, instead of nudging cards one at a time (ratified 2026-08-07, `CLAUDE.md` focus-board block)._

- [2026-10-17.html](docs/gavel-sheets/2026-10-17.html) — **wave 5 (34 lines: 30 Your-Move cards + 3 overnight gates)**, cut by the overnight captain 10-17; JSON + map alongside.
- [render-gavel.mjs](docs/gavel-sheets/render-gavel.mjs) — **the renderer (build sheets with this, never hand-author).** Takes a verdicts JSON → the house HTML (dark/light, print-safe, chip vocabulary green=my call / amber=your judgment / hold=sequenced). Carries its own verifier: numbering 1..N contiguous, every line has question+call+evidence, `<li>` round-trip == input count; exits 2 on any failure. Writes a sidecar `.map.json` (line number → Notion card id) so answers apply back after the session ends.
- [2026-08-15.html](docs/gavel-sheets/2026-08-15.html) — **wave 2 (55 lines from 50 cards).** Every card re-verified against live state by a 10-agent fan-out + 1 adversarial reviewer; **22 of 50 were stale** (7 already done, 14 needed rescoping). Reviewer overturned two money-path calls (the $75.8k Xero void; the phantom-AR/hand-entry sequencing that could drop ~$24.6k) and caught a Klaviyo-only search that would have killed the Birthday flow. Map: `2026-08-15.map.json`.
- [2026-08-07.html](docs/gavel-sheets/2026-08-07.html) — **wave 1, the prototype** (48 lines from 72 cards; cleared 44 in one sitting). Hand-authored HTML — superseded by the renderer above. Still holds the unanswered books/merch/strategy block with calls pre-filled.

## Other (`docs/`)

- [fable-exit-brief-2026-07-07.md](docs/fable-exit-brief-2026-07-07.md) — **Fable's last-day snapshot of judgment** for whichever model drives next: open-arc reads (xero audit first, GG merge gate, corpus > surfaces, Tell gates hold), what makes the stack work, decay clause (`log.md` outranks it on any disagreement). Routed from `docs/rules/wiki-judgment-norms.md`.
- [xero-sync-audit-2026-07-07.md](docs/xero-sync-audit-2026-07-07.md) — **read-only audit of the prod accounting layer (Fable, last-day).** Verdict: sync healthy (34/34 green, current books clean); findings X1–X10 ranked — $115K open WineDirect-era invoices (known-state, see posture banner in xero-accounting.md), reconcile-check page-1 truncation, unbooked $25k loan draw, latent negative-payout bug, token-rail gaps. Read before any shopify-xero-sync hardening.
- [xero-hardening-pr-review-2026-07-11.md](docs/xero-hardening-pr-review-2026-07-11.md) — **cross-vendor review record for the X2/X4/X5/X6 hardening branch** (Fable Tier-1 + Codex Sol blind, reconciled). Verdict: approve after one fix round — C1 fee-journal self-heal lie (required) + 3 ride-alongs; Codex raw findings in appendix. The pre-merge checklist for Kyle's verify sitting lives here.
- [xero-backfill-spec-2026-07-07.md](docs/xero-backfill-spec-2026-07-07.md) — **design spec for booking the 2021→Jan-2026 sales hole (~$5.5–6M) from the frozen wd_* archive.** Era map (2018–2020 good/verify-only · 2022–2025 generate monthly), source tables + quirks, 6 Kyle decision boxes (A–F), 5-step Opus execution plan. The "backfill in its own way" project.
- [2026-10-17-retro.md](docs/retros/2026-10-17-retro.md) — **quarter-3 retro (32-day window; older transcripts rotated off disk).** Marker density 0.06/msg, 46% of user messages are headless `-p` rigs, the 08-07 dep list never ran → `scripts/dep-inventory.sh` writes [dep-inventory-latest.md](docs/retros/dep-inventory-latest.md) (read-only `npm outdated`).
- [2026-07-06-retro.md](docs/retros/2026-07-06-retro.md) — **baseline quarterly session retro** (45d transcript mining: top patterns w/ counts, corrections profile, next-quarter diff targets). New reports land in `docs/retros/`; method = `/retro` skill, nudged quarterly by launchd `com.thief.retro-nudge`.
- [2026-07-17-era-retro.md](docs/retros/2026-07-17-era-retro.md) — **era pass (Fable window)**: answers the baseline's 5 diff targets with the sprint half; findings = meter-management became visible work post-7/13, multi-terminal state skew, the 07-08→07-16 repeated Supabase ask, `/ferry` skill built from the cleared threshold. Next full quarterly Oct 1.
- [2026-07-25-sprint-reconciliation.md](docs/retros/2026-07-25-sprint-reconciliation.md) — **state audit of the 7/24 night shift** (26 transcripts + 13 repos + 195 cards reconciled by a stenographer terminal). Verdict: zero duplicated writes, 2 near-collisions caught by Kyle not the system. Carries the rescued $49k channel-loss finding, the live-lanes v1 post-mortem (mtime read 9 ghosts as live; control cried wolf every start), the board's two-competing-done-signals table, and the approved-but-unclaimed gap. Read before trusting a "nobody is working on this."
- [a-priori-opening-order.md](docs/a-priori-opening-order.md) — distributor opening-order plan. Pre-rep-call.
- [anthropic-support-email-draft.md](docs/anthropic-support-email-draft.md) — drafted email re: ScraperAPI burn from April 23 degradation. Unsent.
- [tool-research-2026-05-16.md](docs/tool-research-2026-05-16.md) — stack-aware research sweep. **Tier 0 item: Anthropic Agent SDK billing pool split effective 2026-06-15** affects Shop Tool wrapper. Tier 1-2 items pending triage.
- [ns-readiness-audit-2026-08-21.md](docs/ns-readiness-audit-2026-08-21.md) — **NEW 2026-08-21, read-only re-audit of Horizon B4** (headless Opus, fired by ef9016c7). Re-verifies the 2026-07-19 `SECOND_STORE_READINESS_AUDIT.md` findings against a month of remediation: 3 closed and proven (GG receiving routing merged to main, stranded-inventory backlog worked, `restock_queue.store` landed 0-NULL), the rest open and two materially larger — reorder engine still gates on aggregate inventory (blind spot now 20× wider) and beer stocking inverted (433 NS-only vs 13 WW-only). Needs Kyle on 3 items: Xero store dimension, NS hours, app behaviour for NS-only stock. Applied fixes from the 08-21 follow-up land in `log.md` § APPLIED 2026-08-21, commits listed there.

## Method / rebuild (`docs/rebuild/`, `docs/method/`)

- [docs/rebuild/RECONCILIATION.md](docs/rebuild/RECONCILIATION.md) — **NEW 2026-06-15.** The CLAUDE.md→v2 layer-separation routing + gated audit (v1→v3): per-item homes, the §16 numbered coverage table, gating rules (verify-then-drop, archive-with-landing), source-completeness checklist. The worked record of the restructure. PROVISIONAL.
- [docs/method/claude-os-conversion.md](docs/method/claude-os-conversion.md) — **NEW 2026-06-15.** The conversion **method** (8 steps, each with its proof) + the **7 axioms** (the through-line the whole OS derives from). Project B's middle leg — the procedure for converting any CLAUDE.md into a layered operating system. PROVISIONAL until validated by living on the Thief restructure.

## Operator's Stack (consultant OS IP — `operator-stack/`, grouped 2026-09-20)

_The "productize the OS" cluster — downstream of the method doc above. Parked; re-enter deliberately, don't dabble._

- [operator-stack/operator-stack-draft.md](operator-stack/operator-stack-draft.md) — the stripped, business-agnostic **public essay** ("The Operator's Stack" — a knowledge architecture for directing AI agents from the orchestrator's seat; `[your domain]` placeholders mark the transferable/non-transferable seam). Draft, not for publishing. Pitch-form sibling of `docs/method/claude-os-conversion.md`. Has the original bit: the method/muscle/person "three honest layers."
- [operator-stack/operator-stack-composition.md](operator-stack/operator-stack-composition.md) — internal scoping map: the system pulled into named parts, each tagged **common / rare / sauce** (the honest "what's actually ours"). Upstream of the public draft. Not for publishing.
- [operator-stack/operator-stack-episode-01.md](operator-stack/operator-stack-episode-01.md) — Episode 1 newsletter **skeleton** (amnesia hook → name Cline → fleet twist → one liftable tactic → soft subscribe). `[KYLE: ...]` = first-person claims only he makes. The content/top-of-funnel layer the Simon Scrapes teardown flagged as missing.

- [drpottsy/winery-pos/README.md](drpottsy/winery-pos/README.md) — **NEW 2026-11-13.** First look at a flat-priced wine-club app for small wineries on Shopify (Commerce7 alternative). The read, open risks, next steps; four research files under `research/`, clickable mockup under `mockup/`, first working release run under `app/` (dev store only).
- [drpottsy/bottle-shop-census/README.md](drpottsy/bottle-shop-census/README.md) — **NEW 2026-10-06.** Which website platform licensed bottle shops run on. NY (3,446 licenses): City Hive 293, Shopify 82, ~80% no findable site; lead CSV of 713 shops. Domain-guess method + its undercount ceiling; Verve positive control. Sizing for a possible bottle-shop software company with Evan.

## Profit_Builder (WRC consulting — separate client work)

- [Profit_Builder/CLIENT_CONTEXT.md](Profit_Builder/CLIENT_CONTEXT.md) — Wildhorse client context: engagement model, systems, people, AI-data posture, project backlog. Read first.
- [Profit_Builder/PROJECT_BRIEF.md](Profit_Builder/PROJECT_BRIEF.md) — Wildhorse Resort & Casino marketing dashboard. Monthly deliverable.
- [Profit_Builder/MONTHLY_WORKFLOW.md](Profit_Builder/MONTHLY_WORKFLOW.md) — data pull → PowerPoint → PDF workflow.
- [Profit_Builder/DASHBOARD_SPEC.md](Profit_Builder/DASHBOARD_SPEC.md) — slide-by-slide build spec.

**`profit-builder-wrc/`** — the Power BI v1 build repo (separate GitHub repo `pottsy17/profit-builder-wrc`, cloned on the PC 2026-08-31; gitignored by the parent like every nested repo). Client data lives in `pbi/data/` and is hard-gitignored, as is `*.pbix` — the saved dashboard embeds the whole dataset. Tracked text only:

- `OPERATIONS.md` — **the runbook**: monthly imports (mail file, DMR, narrative), the runas rule, the nightly job, rules learned the hard way, definitions of record, open items. Read first.
- `ARCHITECTURE_V2.md` — the v2 gameplan (2026-09-07): layer map (box = plumbing, repo = thinking), what goes in `Sandbox`, model changes, backfill sources, deployment path to Diane, order of work, Kyle's gates.
- `WRC_DATA_MODEL.md` — the Oasis data model: access recipe (Mac kinit / Windows `runas /netonly`), every promo-group series decoded, tables that matter, business rules, gotchas, decisions, open questions.
- `sql/01_campaign_views.sql` · `02_redemption_view.sql` · `03_mail_segment_history.sql` · `04_fact_campaign.sql` · `05_fact_campaign_table.sql` — the Sandbox staging layer, in build order (views → the one snapshot table → the fact SQL → its materialized table + refresh proc). Every file carries its control query. 04 is the source of truth; 05 is generated from it.
- `sql/06_casino_metrics_views.sql` — `kp_v_rated_monthly` + `kp_v_casino_metrics_monthly` (19 of the 27 deck KPIs in the CSV's long shape; definitions pinned in comments) · `sql/07_agent_job.sql` — nightly 06:00 SQL Agent job running the refresh proc.
- `pbi/swap_casino_source.py` — repoints `FactCasinoMonthly` at the materialized KPI table (+ CSV for the 6 floor metrics and 2 %-carded ratios); idempotent.
- `dump_fact_agg.ps1` — player-free aggregate of the fact (series × period × segment × predominance × tier) → `pbi/data/fact_agg.csv`, so a terminal can build/control offline.
- `pbi/build_postforma_live.py` — re-points the DM Post Forma Active pages at `FactCampaign` (PF measures, `pf_level`, latest closed campaign via `DimCampaign.is_current`).
- `pbi/build_yoy_live.py` — re-points the *DM Post Forma - YoY* page at the fact (YTD + per-day matrices over `PFYoYRow`×`PFMetric`, hold cards).
- `sql/03b_mail_segment_history_series.sql` — migration: `series` column + wider key so the inactive mail file loads alongside the active one.
- `sql/08_db_yoy_and_new_members.sql` — Mail Program YoY + New Member Stats feeds (long format, mirror the extracts); `pbi/swap_long_sources.py` re-points `DbYoY`/`NewMemberStats` at their materialized tables.
- `parse_dmr.py` · `load_floor.ps1` · `sql/09_floor_monthly.sql` — the six floor KPIs from the month-end DMR PDFs (`Casino Metrics/`, gitignored) → `kp_floor_monthly`; 09 is the full 27-metric KPI view.
- `sql/10_narrative.sql` · `load_narratives.ps1` · `pbi/build_narrative.py` · `narrative_pack.py` — the executive-summary pipeline: month files + old-deck text → `kp_narrative` → live card on the Exec Summary page + the Narrative Archive page; the pack script writes each month's numbers for drafting. Prose + voice file live OUTSIDE git in `C:\Users\kpott\wrc-data\narratives\`.
- `pbi/add_dmr_note.py` — DMR footnote on the four pages that show floor figures.
- `pbi/verify_pbir.py` — deterministic verifier: every visual field reference must exist in the model; run after any page/model build.
- `dump_kpis.ps1` — live KPI table → `pbi/data/kpis_live.csv` (runas window) so `build_report.py` derives the rolling window/floors from live data.
- `load_mail_files.ps1` — loads Crystal "for me" mail files (outside git) into `kp_mail_segment_history`; idempotent, prints file-rows vs loaded-rows.
- `pbi/build_campaign_fact.py` — writes `FactCampaign` + `DimCampaign` TMDL (live from Sandbox), starter measures, relationships; extends `DimDate` with bare MAX (never CALCULATE — cyclic reference).
- `wrc-check.ps1` — LDAP credential check for the non-domain Windows box (separates bad-password from SSPI failure; prompts, stores nothing).
- `pbi/PBI_BUILD_V1.md` — the six-page v1 plan (spec of record; owned by the laptop session).
- `pbi/DAX_MEASURES.md` — DAX source of record: model setup, the six measures, dynamic format string, deck palette, measured gap map.
- `pbi/BUILD_WALKTHROUGH.md` — click-by-click model build, each step ending in an observable **Check:**. Written 2026-08-31 after "Kyle is PBI-fluent" was measured false.
- `pbi/measures-to-paste.txt` — the six measures as fence-free plain text (Markdown code fences were being pasted into Power BI).
- `pbi/build_postforma_pages.py` — deck pages 9–12 (Matrix, Post Forma Active/YoY/Inactive) + Scorecard, Trend Explorer, Mail-by-Market pages; generates the four Post Forma TMDL tables (2026-09-03).
- `pbi/extract_postforma.py` — tidies the Mail Post Forma `.xlsm` (levels by market, YoY long+fmt, hold %, matrix tiers), totals controlled against the May spec (2026-09-03).
- `pbi/build_report.py` — report-as-code: writes PBIR visual JSON + TMDL measures into the `.pbip` (exec summary, DB YoY page, New Members page, filter repair), schema-validated (2026-09-03).
- `pbi/extract_tables.py` — tidies the two staging dumps (DB YoY, New Member Stats) into long CSVs for the page-5/6 matrices; self-checked, YTD sums verified against the deck spec (2026-09-02).
- `pbi/PAGE_BUILD_CLICKS.md` — the click layer for the sheets: Format-pane anatomy, per-series colours/widths, page-level year filter, copy-a-chart loop, footer, page-nav instead of bookmarks, last-point-label DAX (2026-09-02).
- `pbi/PAGE_BUILD_SHEETS.md` — pages 2–6 as copy-the-chart recipes, plus the **Y-axis ruling** (Kyle 2026-09-01: no forced zero baseline, 5-year trailing window) and the per-page gap/zero rules.
- `pbi/kpi_pack.py` — generates the May-cycle exec-summary KPI pack; mirrors the DAX exactly (percent AVERAGE, missing month BLANK), four self-checks. `py -3 pbi/kpi_pack.py`.
- `pbi/anomaly_scan.py` — the true-up checklist: integrity, coverage gaps, real zeros, per-metric robust-z outliers, suppressed by a gitignored `events.csv` that records why a period is unusual. `py -3 pbi/anomaly_scan.py`.

## Personal / legacy

- [lieu-dit/README.md](lieu-dit/README.md) — **NEW 2026-10-01.** Lieu-Dit, the daily wine map game (own repo `~/code/lieu-dit/`, gitignored here). State, what-talks-to-what, run commands, shape licences, open list. Ownership undecided.
- [sulfite-club/README.md](sulfite-club/README.md) — **NEW 2026-09-19.** Kyle's own wine-shirt brand, **not a Thief project** (own repo `pottsy17/sulfite-club`, gitignored here, arm's-length from the shop on purpose). Python drawing toolkit → 17 outlined one-ink print files, byte-identical on the PC and the Mac. Siblings: `IDEAS.md` (the group's whole joke list, verbatim, with status + flags) · `BRIEF_URL.txt` → the brief page (verified fees, margins, clearance, Kyle's 7 moves). This pointer is the whole wiki footprint; detail stays in that repo.
- [frenchfarmersonly/README.md](frenchfarmersonly/README.md) — **NEW 2026-08-29.** The ceremony joke made real: password-gated wedding photo/video share for Morgane & Christophe on Cloudflare Workers + R2 (personal kpottorff account, $0). Stack, secrets, corner-cut ledger, and the drop-in slots (Spring 1 music, intro video, profile photos) all in the README. Live: `frenchfarmersonly.kpottorff-2cb.workers.dev` → frenchfarmersonly.com once NetSol nameservers repoint.
- [bowie-movie/README.md](bowie-movie/README.md) — **NEW 2026-09-26.** Kyle's personal screenplay idea, not a Thief project: a "reverse Terminator" where 2026 goes back to save Bowie. `research/` (5 files), `OUTLINES.md` (4 outlines), `COLD_OPEN.md`.
- [dad-stories/README.md](dad-stories/README.md) — **NEW 2026-08-16.** Dad's life stories, StoryWorth replaced by a mic: Lark M2 → Voice Memos → `transcribe.sh` (whisper medium.en, $0) → Claude shapes `stories/` + next visit's 3 questions. Mom next.
- [carry-the-light/carry_the_light_context.md](carry-the-light/carry_the_light_context.md) — children's book context doc.
- [carry-the-light/manuscript/carry_the_light.md](carry-the-light/manuscript/carry_the_light.md) — manuscript, current.
- [carry-the-light/manuscript/EDITORIAL_REVIEW.md](carry-the-light/manuscript/EDITORIAL_REVIEW.md) — voice-pass review.

## Tasting menus (`tastings/`)
- `tastings/themes-2026-10-29-to-12-17.md` — Thursday theme PROPOSAL for the six open OND Thursdays, 2 candidates each scored against 315 `thursday_tastings` rows + the metrics view, my calls pre-filled; BMAC 11-05 is the open question. Born 2026-10-11 (OND Lane 1).

- [tastings/private/tasting-menu-format.md](tastings/private/tasting-menu-format.md) — private-tasting **bifold format spec**; the template `private-tasting-menu-bifold.html` sits beside it. (Moved from root 2026-09-20; the May 2026 Indiana wine-bar-owners menu is in the same folder.)
- [tastings/archive/README.md](tastings/archive/README.md) — **NEW 2026-08-14.** The `.docx` menu archive → Supabase pipeline (2019-2026): 294 events + 1,975 pours + 3 views, deterministic/$0. Read before querying tasting history or re-running the loader — it names the traps (`name_key` must be one implementation across both eras; repeat view excludes bonus pours on purpose; booklet print-order; filename beats the printed date; a backfill must stamp `blog_published_at` or arm a 249-post blog spam). **+ §The Thursday Book (2026-09-26):** the archive read back against the shelf + the till — 320 in-stock wines with an unused house note, the searchable 8-year book, 935 Thursday regulars; `thursday-book.mjs` → private artifact https://claude.ai/artifact/BytfXSJFKwvcnfQ5MoGKDj.

## Toys (root — deliberately not lanes)

- [thief-wrapped/README.md](thief-wrapped/README.md) — **"The Thief: The Story So Far"** (2026-07-18, Fable free-rein night). The shop's 8-year sales history as a story page — self-contained `index.html`, private artifact (URL in log 2026-07-18), provenance + refresh queries in the README. Snapshot, not infrastructure; any customer-facing future = fork conversation first. Sibling toy: the Daily Tell (`Blind-Tasting-App/tools/daily-tell/`).

## Memory (lives outside this repo)

- Cross-conversation memory → `~/.claude/projects/-Users-pottsy-code/memory/` (self-indexed in its own `MEMORY.md`). The empty `~/code/memory/` stub was abandoned cruft, removed 2026-06-15.

---

## Notes on this index

- **What's in here:** project state docs, plans, specs, rule files, vision docs.
- **What's NOT in here:** knowledge-base content files (grape guides), manuscript versions, individual social post drafts, marketing collateral copy. Those are content, not project state.
- **How to use:** start any session by checking this file. If a project isn't in here, it doesn't exist as ongoing work.
- **How to maintain:** when a new .md file is added, add a one-line summary here. When a doc is spent or superseded, `git mv` it to `docs/archive/YYYY-MM/` and collapse its entry into the "Archived" line below (the groundskeeper's weekly report lists candidates).

## Civic (`civic/` — Kyle's policy ideas, not Thief business)

- [civic/MAIN_STREET_ACTIVATION.md](civic/MAIN_STREET_ACTIVATION.md) — downtown vacancy property-tax concept for the state legislators; concept + next steps.

## Archived (`docs/archive/YYYY-MM/` — git-tracked, hidden from search by `~/code/.ignore`)

_Spent or superseded docs move here with `git mv`; nothing is deleted. One line each. To search the archive: `rg --no-ignore <term> docs/archive/`. (`_archive/` at root is a different thing: gitignored old code prototypes.)_

- **2026-09-20** — `SPRINT_FINALE_2026-07.md` (Fable-finale sprint plan, closed 7/19) · `MORNING_2026-07-11.md` (Saturday unstick list) · `PARKED_FOR_OPUS.md` (post-Fable queue; regen cancelled 7/18, PRs applied 7/16) · `LANE3_EXTRACTION_STRATEGY.md` (tabled 5/30 LLM-vs-deterministic debate; corpus built 7/11) · `CLAUDE.v1-backup.md` (pre-restructure CLAUDE.md, 6/15). · `blind-tasting-ios-SETUP.md` (March Xcode setup for The Tell's first screen; the code moved to `TheTell/` and every "next sprint" item on it is built — folder removed). · `WWDC26-notes.md` (June session digests; what mattered is in `docs/rules/ios-apps.md`). · `ONBOARDING.md` (May 2026 whole-stack map, written for a senior dev + used to brief ChatGPT and claude.ai at the time; a snapshot — regenerate fresh rather than trust it; the paste-first passport for outside AIs is now `fable-kit/COMPACT_CONTEXT.md`).

## NS migration (`_ns-migration/`)
- [_ns-migration/README.md](_ns-migration/README.md) — **audit trail of WW→NS inventory migrations**, one wave per section + per-item CSVs/logs. Wave 1 (2026-08-07): all `Discount:4-Pack` beer — 461 moved (4,859 units) + 201 re-homed; open tails: 17 non-beer review items, 29 oversold rows. Runner: `golden-goose/review-ui/scripts/migrate-4pack-beer-to-ns.ts`.

## Beer planogram (`beer-planogram/`)
- [beer-planogram/README.md](beer-planogram/README.md) — **NS beer-fridge schematic (started 2026-08-10)**, built from combined-store sales history (one beer program; packaged beer moved WW→NS in Wave 1). `walk-sheet-v1.html` = draft door/shelf map awaiting Kyle's red-pen → v2 laminated staff card. Parked on purpose: 25% mix-4 pricing rethink.

## Bank reconciliation (`bank-rec/`)
- `bank-rec/BRIEF_XERO_COMPARE.md` — execution brief: bank-archive ↔ Xero comparison + gated gap-fill (the check coding engine)
- `bank-rec/compare/SUMMARY.md` — **Phase A result (2026-07-19, read-only)**: 7,647 lines / $18.2M classified; what's already in Xero, what needs an entry, what needs Kyle
- `bank-rec/compare/CODEX_REVIEW_2026-07-19.md` — Codex implementation review of the Phase B write rail, both rounds + 4 accepted residual risks
- `bank-rec/compare/FABLE_DECISION_REVIEW_2026-07-19.md` — **Fable decision review of the Phase B rail: CONDITIONAL GO, 3 gates** (A: rehearsed unwind + regenerate before canary · B: accountant sitting before filed years 2021-24 · C: shrink the 9999 pile). The review of record for whether to post at all.
- `bank-rec/ocr/CONTROL_NOTES.md` — OCR corpus integrity: two-tier read quality, contaminated batches, memo-field version boundary
- `marketing/klaviyo/drop-recovery-email-draft.md` — drop-recovery email draft (one note per past WD buyer listing their unsold wines); awaiting Kyle go (2026-09-04)
- `marketing/klaviyo/F2S_RUNBOOK.md` — **NEW 2026-09-25.** Kyle's 30-min sitting: the +21-day first-to-second step inside live `S39TcX` (Step 0 = fix the "since starting this flow" filter that has mailed repeat buyers since 08-03), Collections split, 10% holdout stamped `f2s_group`. Facts in `f2s-event-fields.md`; baseline 5.8% at 21 days.
- `marketing/ads/README.md` — **NEW 2026-09-25.** The paid-lanes prep pack (Google Search · Meta local · Seattle · Tri-Cities · billboard fork): lane map, `PIXEL_SITTING_RUNBOOK.md` (Kyle's 45 min), keyword + ad CSVs with `check-ad-copy.mjs` verifier, `meta-audiences.md`, `creative-shot-list.md`, two landing-page drafts, `utm-convention.md`, `billboard-quote-request.md`. Nothing live; zero spend before the test order.
- `marketing/winback/2026-09-25-winback-champagne-01-draft.md` — **NEW 2026-09-25, RATIFIED.** Winback email #1 for the Champagne branch (114 lapsed mid, their houses, the Oct 22 tasting hook); the other region branches follow this shape.

## Added 2026-09-26 — removal test + verdict queue
- `docs/rules/removal-test.md` — the two-week prune experiment: what moved where, the meter, the ruling rule.
- `docs/rules/landmines.md` — SPINE §3 full text (verbatim move), §1 ladder story, §6 smaller-models note.
- `docs/rules/projects.md` — every repo + purpose + the nested-by-design folder layout (verbatim move from CLAUDE.md).
- `docs/reviews/2026-09-26-codex-system-review.md` — Codex (gpt-6-astra, low) blind review of the operating system + the brief it got.
- `docs/archive/removal-test-2026-09-26/` — pre-prune CLAUDE.archived.md / SPINE.md / MEMORY.md, the rollback (renamed 2026-10-03 so it no longer loads as a nested CLAUDE.md).
- `.verdicts/` + `scripts/verdict.sh` + `~/.claude/commands/verdicts.md` — the shared bounded decision queue.
- `docs/reviews/2026-09-26-ecc-steal-list.md` — ECC read on disk: three mechanisms worth stealing (instinct ladder, context-size compaction, harness tests), what we already have, what to skip.
- `dreaming/instinct.mjs` + `dreaming/instincts.jsonl` + `dreaming/instinct.test.mjs` + `dreaming/seed-2026-09-26.sh` — the instinct ladder: deterministic confidence score + altitude gate over every pattern dream has seen (`docs/rules/dreaming.md` §Instinct ladder). `dreaming/runs.log` = durable record of each timer fire.
- `harness/hooks/scratchpad-import-guard.mjs` + `harness/hooks/tests/scratchpad-import-guard.test.mjs` — PreToolUse Bash guard for unresolvable Node imports (the ledger's first FAILED RULE, mechanized 2026-09-26); the first test suite on a harness hook, including a run through a symlink.
- `docs/reviews/2026-09-26-codex-readonly-autoallow-v1-hold.md` · `-v2-hold.md` · `-v3-hold.md` — three Codex holds on a hook that auto-approved read-only commands; the hook is RETIRED (`harness/README.md` §Retired). The rule that came out of it: a hook may deny or warn, never approve.
- `docs/reviews/2026-10-10-fringe-audit/REPORT.md` — "Thin Air": what the stack is missing that no list carried (fundamentals + unsaved-for money), 9 lenses × Opus skeptics × critic round; 27 survivors adjudicated with my calls, 8 phone asks, expiry-calendar seed rows, the 6 kills. Same folder: `BRIEF.md` (the lens brief + stay-out/already-listed lists), `findings-*.json`, `_skeptic-*.txt`, `result.json`, `thin-air.html` (the phone decision surface, published https://claude.ai/artifact/1utvoAw99Zkwb9hWSyRpGT).
- `docs/sprints/2026-10-10-ond-sprint.md` — the OND sprint week (10-10 → 10-18): Kyle's 8-item unlock list, one row per lane (seat · status · verifier · next gate), stay-out list, day log. Plan of record in `~/.claude/plans/`. Born 2026-10-10.
- `docs/sprints/2026-10-16-paradigm-audit.md` — which cornerstones were built on a platform that moved (MCP kill rule, Context7 auth, `-p` rail, fleets vs Workflow, 65 crons vs Routines, fable-kit, Brave): 13 rows, my call + verifier each, order of work. Born 2026-10-16.
- `docs/reviews/2026-10-01-exposure-audit/REPORT.md` — exposure audit of the main Supabase project (what the internet and any customer account can reach): 35 confirmed holes, worst first, with fix drafts under `fix-drafts/`. Same folder: `BRIEF.md` (the fleet's one brief, every role) · `coverage.mjs` (the deterministic verifier; re-runnable) · `workflow.js` + `wait-test.mjs` (the workflow that waits out a closed usage window, and its test). · `ROTATION-SITTING.md` (Kyle's key-rotation runbook for rows 3-6, 2026-10-02).
- `dreaming/harness/2026-09-26-guard/replay-corpus.mjs` — replays every distinct `execute_sql` from recent transcripts through the live guard and a candidate, prints the transition matrix against what Postgres actually said. Run before ANY edit to `sql-schema-guard.mjs`.
- `marketing/wine-alliance-ad/README.md` — Wine Alliance booklet full-page ad (Dec 2026 print): build.py → outlined-vector CMYK press PDF; copy rulings.
