Skip to content

Session handover — 2026-07-14 pt3

A follow-on to pt2. ADR-0011 Slices 1c and 1d shipped; Slice 2 (the VLM) re-asked per the ADR and NOT built. The Slice-1 program — "make the gate see what the pipeline already measures" — is complete. Two commits: 5460cfd (1c), fb2400f (1d). WCP worker b7b37455 deployed.

Where the two sites now stand

site verdict why
WCP HOLD contentDestroyed (the gallery intro — schema has no home for it) + videoDebt (live references 27 videos, we serve 1). Composite 87.2 — above τ_ship; only real, named work holds it. S 82.2 / T 100 / G_det 84.4.
garvanbay HOLD criticalSectionDropped (/contact has no hero) — unchanged, untouched all session (seed A/B: 0 diffs; worker not redeployed).

The three systematic tone vetoes from 1a are gone for real reasons: variant mis-renders 38 → 0 on the re-measurement.

Slice 1c — chrome + assets now gate (5460cfd)

See the ADR-0011 tracker row for the full account. The headlines, both measured before building:

  • The "7/7 chrome dims agree" was FABRICATED, not un-compared — topbar.present scored two blind probes' false === false as agreement (live: zero-height pinned container; ours: a Wix-only selector). Capture schemaVersion 6→7 (topbarStrip measures the strip on BOTH lanes, content.bgImages, page-level mediaRefs); globals gained a palette tier; positive chrome disagreements now veto (chromeDisagree). See known-patterns "A presence dim computed from two blind probes FABRICATES agreement".
  • asset-manifest was never promotable (written PRE-migration — it reports the SHIP reference as 63 raw refs; and its expected side is a lazy-mount floor reading 0 videos on a site that plays video). The veto basis is the rendered-DOM mediaRefs harvest instead: live WCP references 27 of the pool's 46 videos (19 = library cruft — the split that was never done); ours serves 1 → videoDebt HOLDs with a per-page id punch-list. Plus deadAssetRef from the post-migration gaps sidecar. See known-patterns "The 'used on live' asset census is the RENDERED DOM" and known-issues "asset-manifest.json is a PRE-migration artifact".
  • imagesDropped measured and NOT promoted: the bgImages recompute took garvanbay 34 → 14 — better, not ~0, so it stays reporting-only (residual = aligner anchor drift).

Slice 1d — the lever audit (fb2400f)

  • The instrument: apps/cms/scripts/lever-audit.mjs + DEV-only /lever-audit page; matrix derived from the schema (_definition), computed paint/layout signatures only. First run convicted six levers: four real, in three mechanisms (inline style beats tone; adapter drops the field ×2; scoped unlayered block CSS beats layered utilities) — all fixed, allowlist empty — and two were the audit's own blind spots, fixed in the instrument. See known-patterns "The lever audit".
  • The FAQ navy (19 pages, the largest G_det item) was the inversion of the ribbon: a WORKING lever fed a WRONG value by a dishonest join key — one Wix wrapper id spans the cta ribbon and the FAQ band; the live FAQ's id doesn't exist in the static DOM. Fixed with a heading/textHead consistency guard on makeStyleLookup (A/B: garvanbay 0 diffs, WCP exactly 19 FAQs); 19 re-stamped via sync-style-to-d1 + committed seed. See known-patterns "A comp-id join is only as honest as Wix's WRAPPER STRUCTURE".
  • cta-wcp paints a SET tone via Section's CVA (solid brand — faithful to live's solid #1e73be and probe-visible); UNSET keeps the legacy gradient (unmeasured must not change).

Slice 2 re-ask — NOT built (recorded in the ADR tracker)

Every item on the owner's 4/10 list now has a deterministic detector: topbar → chromeDisagree, FAQ colour → fixed + systematic machinery, icon cards → contentDestroyed, nav links → fabricatedLink, videos → videoDebt. The remaining HOLDs are render/schema work, not perception gaps. Reopen trigger unchanged: a defect mapping to NO existing dimension.

Filed, not fixed (with triggers, in known-issues)

  • Home fcr.faq heading is cta copy — heading-EXTRACTION matcher bug; blocks the 20th FAQ tone fix and causes home's faq dropped S row.
  • Probe gradient-parsing — moot on both references post-fix; trigger = a site whose LIVE bands are gradient-painted.
  • apps/cms/src/components/blocks-v3/ is an orphaned pre-workspace copy, zero importers (chore).
  • garvanbay's worker carries the pre-1d bundle; the canonical changes reach it on its next deploy (provably no-op there: seed A/B 0 diffs, default/unset tones render byte-equivalent CSS).

Next

  1. The fcr.gallery body field (agreed next slice) — matcher extraction → schema → transformer → adapter → block; closes WCP's contentDestroyed veto. Needs seed regen through the graft check (the copy isn't in D1 to patch).
  2. Video segmentation/render workstream — the 26 missing videos; videoDebt keeps it verdict-visible; pulling files without section homes is motion without progress.
  3. Still open from earlier: garvanbay /contact hero (its HOLD), the live-asset harvest's wider use, aligner role-aware anchoring, Slice 6 calibration.

The through-line

pt2's was check whether the instrument's finding ever reaches the verdict. This session's is the completion of that arc, twice over: 1c — two instruments that "agreed" were agreeing in a shared blind spot, and the honest census had to be built before the veto could be; 1d — the audit's first six convictions included two of its own blind spots and one case where the lever worked and the data lied. Same discipline every time: measure the instrument before believing it, and let the negative controls (garvanbay, 0 diffs, three times) prove the fix touched only what it claimed.