`5x/REPORT.md`

42 lines
# 5x REPORT — dw-pairs-well CLIP build (local :9821, PAIRS_CLIP_DEFAULT=1)

Target: http://127.0.0.1:9821/preview.html · visual-search :9914 (launchd, 334k vectors)
Build under test: commits 52301ea → 7fcb7a9 (CLIP similar engine, CLIP pairs blend,
ACTIVE+online_store_published link guarantee) + ce3fda2 (circuit breaker, ext. session).

| sweep | verify | caught | fixed | commit |
|-------|--------|--------|-------|--------|
| 1 | 6/7 (six-way all PASS; Phase-C 0/10) | 1 — shared 3x clickthrough.js is hardcoded for the wallco curator UI (selector mismatch → all timeouts). Six real layers had 0 errors. | project-specific 5x/clickthrough.mjs exercising the REAL controls; found + encoded the lazy CTA-expand flow | 8b2123a |
| 2 | 10/10 (6-way + own clickthrough 6/6) | 0 | — | — |
| 3 | 12/12 (incl. real Safari + Firefox) | 0 | — | — |

**Stop condition met: 2 consecutive clean sweeps (2 & 3).**

## Final six-way state
M1 HTTP 200 · M2 headless render OK · M3 E2E 0 JS errors ·
B4 Chrome PASS · B5 Safari PASS · B6 Firefox PASS ·
Phase C 6/6 (chip→CTA→grid populate, /products/-shape links, #sku+Load, paint toggles,
grid swap) · 0 console/page errors across all sweeps.

## Adjacent evidence (same build, earlier this session)
- tools/live-link-check.sh: 128/128 suggested links (8 heroes × pairs+similar, CLIP on)
  → HTTP 200 on the LIVE store. 0 broken.
- CLIP quality A/B: damask hero → same-pattern-other-colorway + theme matches across
  5 vendors (vs one-vendor tag grab-bag); grasscloth hero pairs → 6/6 style-world hits.

## Open items (not defects — gated on Steve)
1. Prod deploy of this build (classifier-gated): `bash ~/Projects/_shared/scripts/deploy.sh`
2. CLIP service exposure for Kamatera (tailnet serve or run-on-Kamatera) — until then
   live serves the tag engine with the same link guarantee.

## Post-contrarian sweeps (FIX FIRST applied)
| sweep | verify | caught | fixed |
|-------|--------|--------|-------|
| 4 | 4/4 six-way; Phase C 5/7 | no-store applied (contrarian fix). 2 Phase-C fails: (a) panel collapses on source change — test must re-expand; (b) designed 404 console logging tripped harness | test-flow fixes + color+texture lock + PUB_SQL boot probe (0d36db3) |
| 5 | 4/4 six-way + 7/7 Phase C, 0 errors | 0 | — |

Cap of 5 reached with sweep 5 fully clean. Honest note: only ONE clean sweep follows the
sweep-5 code changes (color lock, PUB_SQL) — the strict clean-twice bar was met for the
pre-contrarian build (sweeps 2+3), and post-fix by sweep 5 alone. Live link-check re-run
against prod after deploy is the compensating evidence.