`scripts/buyability-audit/FINDING.md`

51 lines
# Buyability = HEALTHY (0 active-but-unbuyable products — a full CENSUS, not a sample) — clean bill for the buyability question only (NOT a fleet all-clear); inventory=2026 drift splits into two populations; 73.9% CONTINUE-oversell is a named downstream residual — cycle 89

> **Narration correction (officer):** the 100% buyable result is a full CENSUS of all 71,719 active products (Shopify sets `availableForSale=false` only when DENY+qty≤0 or unavailable, so the census IS the structural guarantee). The 3,000-product probe is a flag-HONESTY audit (belt-and-suspenders), NOT the proof — don't read it as "4% sample proves 71,719." And this is a clean bill for the BUYABILITY question only, not a fleet health all-clear.

**DTD verdict B (unanimous 3/3 — Claude + Codex + Qwen).** READ-ONLY (Admin GraphQL full active scan + a 3,000-product policy-mechanism probe; no mutation), **$0**. Audits the sharpest customer-facing inventory defect — an ACTIVE product NO shopper can purchase (every variant out-of-stock + inventoryPolicy=DENY = a dead-end "can't add to cart" PDP) — and measures drift from the standing inventory=2026 both-variants convention.

## 🟢 Result: ZERO active-but-unbuyable products — every one of 71,719 active products is purchasable
Full active scan (71,719, complete=true, PJ excluded):
- **BUYABLE (≥1 variant availableForSale): 71,719 (100.00%)**
- **UNBUYABLE (no variant purchasable — the dead-end defect): 0 (0.00%)**
- no-variants data oddity: 0.
→ No active product blocks the sale. The c88 no-image case (orderable-but-blank) has no buyability analogue — nothing is active-but-unbuyable.

## 🔬 Skeptical decode — the census is the proof; a mechanism probe AUDITS that the flag isn't lying (don't trust availableForSale at face value)
The full census already guarantees buyability (availableForSale=false only on DENY+empty/unavailable). To confirm the flag isn't lying, I probed the underlying inventory mechanism (3,000 products / 5,095 variants — a flag-honesty audit, not the proof of the 71,719):
- **inventoryPolicy CONTINUE (oversell-enabled, always buyable regardless of stock): 73.9%** — DW's made-to-order/trade model: most variants oversell, so they're purchasable even at 0 stock.
- **inventoryPolicy DENY: 26.1% — and ALL sampled DENY variants have qty > 0** (buyable via real positive stock, incl. the 2026 sentinel).
- **DENY + qty ≤ 0 (the ONLY state that blocks a sale): 0 in the sample.** And 0 "DENY + qty≤0 + availableForSale=true" oddities.
→ So 100% buyable is REAL and structurally explained: 3/4 oversell + 1/4 stocked, and the blocking state (DENY+empty) does not occur. The negative control holds (a buyable bulk of 100% makes any unbuyable product a true anomaly — there are none).

## 📋 inventory=2026 convention: ~79% followed — drift is benign FOR BUYABILITY but splits into TWO populations (officer)
- Active products with ≥1 variant at qty 2026: **56,514 (78.80%)**. Of 27,179 multi-variant products, **23,265 (85.6%) have BOTH variants at 2026**.
- **The ~21% NOT at 2026 are still 100% buyable — but that is ONE label over TWO different populations (officer):**
  - **(a) buyable via positive REAL stock that simply never got the 2026 sentinel** → benign onboarding drift.
  - **(b) buyable ONLY because inventoryPolicy=CONTINUE masks qty ≤ 0** → these sell today but are riding oversell, and the missing 2026 marker is exactly the signal they were never properly onboarded → a SOFT defect, not benign.
- So "benign, do not mass-touch" is correct as a NO-MASS-WRITE call (the default-variant-consume landmine is real), but the (b) subset deserves a follow-up READ (segment it) — it is not a non-issue.

## Disposition: clean bill on the buyability QUESTION; no fix needed (not a fleet all-clear)
1. **Buyability is HEALTHY fleet-wide.** 0 active-but-unbuyable products (full census); the flag-honesty probe confirms the blocking state (DENY+empty) doesn't occur. No fix.
2. **inventory=2026 drift (~21%) is benign FOR BUYABILITY but two-population** (see above): (a) real-stock-no-sentinel = benign; (b) CONTINUE-masked-qty≤0 = a soft onboarding defect. NO-MASS-WRITE (buyable + default-variant-consume landmine `shopify-default-variant-consumes-sample`), but segment + read the (b) subset next.
3. **Together with c86/c87/c88 this extends the storefront-health picture:** discoverability clean (c86/c87), image-presence near-clean (c88, 2 defects), buyability clean (c89). The catalog's active set is structurally sound for browse → PDP → purchase. The April ARCHIVE remains the only material dark/lost-inventory vector (separate campaign).

## What this does NOT cover (named residuals)
- **CONTINUE-masked vs true-stock buyability (officer — NOT a one-liner):** 73.9% of variants are CONTINUE (oversell-by-default). For made-to-order/cut-to-order wallcovering that's often CORRECT (DW orders from the vendor, holds no roll stock), but it means a chunk of the "100% buyable" is "buyable because we never say no" — converting a stock problem into a downstream FULFILLMENT / lead-time / refund problem. Worth its own read: correlate CONTINUE-default against vendor lead-time accuracy. Out of scope for buyability, but real — not harmless.
- **Cart-to-checkout completion** — add-to-cart-able ≠ checkout-completable; measured variant purchasability, not the full funnel.
- **Price-zero free-checkout** (c67/zero-dollar-canary class) — buyable AND mispriced; already covered by the zero-dollar-orderable canary; not re-run.
- **Image CORRECTNESS** (c88 higher-severity residual) + **full-404 pass** + **merchandising-DEPTH** (c87) — still open.
- **inventory=2026 (b)-subset** — the CONTINUE-masked-qty≤0 slice of the 21% drift (segment it next).

_Cycle 89. DTD-picked B (unanimous 3/3). Buyability audit: full active CENSUS (71,719) → 0 active-but-unbuyable products (100% have ≥1 purchasable variant; availableForSale=false only on DENY+empty, so the census IS the guarantee). A 3,000-product probe AUDITS flag-honesty (not the proof): 73.9% CONTINUE oversell + 26.1% DENY-but-all-qty>0, 0 DENY+empty → no blocking state occurs. inventory=2026 ~79% followed (85.6% multi-variant at both); the ~21% drift is benign FOR BUYABILITY but TWO populations — (a) real-stock-no-sentinel benign + (b) CONTINUE-masked-qty≤0 = a soft onboarding defect to segment next. Clean bill for the buyability question only (NOT a fleet all-clear); no fix, NO-MASS-WRITE. Named residuals incl. officer additions: CONTINUE-masked vs true-stock buyability (fulfillment/lead-time risk, 73.9% oversell), the (b)-subset, cart-to-checkout, price-zero (separate canary), image-correctness, full-404, merchandising-depth. PJ excluded. Read-only, $0, no fix gated._

---
## 🏛️ OFFICER SIGN-OFF — vp-dw-commerce: CONFIRMED (narration sharpened; 2 residuals the conclusion under-weighted)
- **Q1 — clean bill SOUND, mildly mis-narrated (fixed):** the 100% result is a full CENSUS (Shopify sets availableForSale=false only on DENY+qty≤0 or unavailable, so the census IS the structural guarantee), NOT a 4%-sample inference. The 3K probe verifies the flag isn't lying (belt-and-suspenders) — reframed so a later challenge to the 4% is answered by "we scanned 100%."
- **Q2 — the 21% drift is TWO populations (fixed):** (a) real-stock-no-sentinel = benign; (b) buyable ONLY via CONTINUE masking qty≤0 = a soft onboarding defect (the missing 2026 marker is the never-properly-onboarded signal). "Benign, do not touch" is right as a NO-MASS-WRITE call, wrong as a single-population label → logged as two-population; (b) gets a follow-up read.
- **Q3 — 73.9% CONTINUE is NOT harmless (promoted from a one-liner to a named residual):** three-quarters of the fleet oversells-by-default. Often correct for made-to-order wallcovering (no roll stock held), but it means part of "100% buyable" is "buyable because we never say no" — converting a stock problem into a downstream fulfillment/lead-time/refund problem. Named explicitly + a follow-up to correlate CONTINUE-default vs vendor lead-time accuracy.
- **Q4 — 3 overstatements fixed:** "verified by mechanism not the flag" (census is the proof, probe is flag-audit); "BENIGN 21%" (benign for buyability, not for onboarding hygiene); "clean bill no fix" (true for THIS audit's question, NOT a fleet all-clear).
- **Disposition:** pass; no prod writes (gated). Recommended cycle-90 follow-up: read-only segment the 21% into "no-sentinel-but-real-stock" vs "buyable-only-via-CONTINUE," + a canary correlating CONTINUE-default with vendor lead-time. Nothing changes the no-fix verdict; it sharpens narration + names two under-weighted residuals.

CONFIRMED. Buyability clean by full census (0 active-but-unbuyable); the no-fix verdict stands for the buyability question; the CONTINUE-mask + inventory=2026-(b) subset are the two named residuals to read next. Read-only, $0.