`data/remaining-cohort-audit-2026-09-02.md`

14 lines
# TK-10978 — Pre-approval audit of the remaining unbuyable dry-run cohorts (2026-09-02)

READ-ONLY re-verification against the current live mirror. **Purpose: stop a stale/broken dry-run plan from being approved.** Triggered by finding that Rebel Walls' claimed identity join yielded 0 in reality.

| Cohort | Dry-run claim | Reality today | Verdict |
|---|---|---|---|
| **Rebel Walls** | "494 clean, cost staged" | RS-series mfr **absent** from catalog; identity join = 0. Content-match: **280 safe** / 10 ambiguous / 204 need re-scrape | Corrected plan drafted → `pending-approval/2026-09-02-TK10978-rebelwalls-494-CORRECTED-plan.md` |
| **Coordonné** | "136 clean, mfr_sku identity join" | Live mfr_sku is dirty (`"6461256 Durable Vinyl"`) → identity join = **0**; numeric-prefix match spurious; naive content-match = **0** (catalog `pattern_name` bakes color: "Leaf Craze White") | **NOT approvable as-is.** Needs real identity reconciliation or a re-scrape. Do not approve the 08-31 dry-run. |
| **Donghia** | "152 clean, mfr_sku identity join" | Mirror shows only **2** unbuyable now | Stale/moot — recovered elsewhere or flags flipped. No action. |

**Cross-cutting lesson (reinforces the methodology memory):** a dry-run plan's "clean count" is only trustworthy if the *join it claims* is re-verified against current data — live `mfr_sku` values drift (dirty suffixes, series changes) and the mirror's `has_product_variant` is unreliable. Always re-probe the join before approving.

**Net residual for TK-10978 recovery:** Rebel Walls 280 (drafted, gated, pilot-first) is the only remaining data-recoverable cohort. Coordonné needs a scoped reconcile project; Track B (Koroseal/Maya/etc.) still needs Steve cost lists (TK-11042); Innovations 51 still needs the vendor recipient + email approval (TK-11041). No writes fired.