Skip to content

Scope — clean-room rebuild of both reference sites + a ranked gap report

Status: WCP EXECUTED 2026-07-13 pt4 (rebuilt from inception, deployed, scored — worker 51891396). garvanbay NOT rebuilt; the live-asset harvest (§5) NOT run. See the pt4 handover for results and the ranked gaps. Preconditions all landed pt3/pt4.

Correction (2026-07-13 pt4): the Risks section's claim that "the rebuild MUST re-use [the authoritative basename→R2-URL] map" is FALSE. The map was never needed. The R2 key is sha256(bytes)-derived, so it is reproducible from the bytes — and a clean-room rebuild re-crawls the bytes. transform-seed-images.mjs hashed them, derived the same keys, HEAD-probed R2 and reported idempotent: 299 — i.e. it recomputed 299 existing keys exactly, uploaded only what was genuinely new, and lost no image.

Why the original was drawn: known-patterns is right that "the R2 key is a lossy hash you can't recompute" — from a basename. The 2026-07-08 session was resolving a regenerated seed without re-running the migration, so a basename was all it had. That premise does not hold for a rebuild. The map is the safety net (for a ref whose local file the crawl never fetched and whose wixstatic source is dead) and an audit instrument — not the path. It was graduated to a committed per-site helper (apps/cms/scripts/r2-basename-map.mjs) anyway, since known-patterns asked for it; it did no work in the rebuild.

Building it also reproduced this repo's signature failure once more: keyed by domain (garvanbay.ie/) it returned a confident zero objects, because the R2 prefix is the site (garvanbay/). An empty enumeration reads as success. Corrected: garvanbay 31 objects, WCP 144, 100% carrying x-amz-meta-original-name. Related: ADR-0009 (measured near-match — the fidelity bar this measures against), ADR-0010 (Wix API content/media plane), ADR-0008 (per-site fleet), ADR-0005 (image pipeline).

Why

Neither reference site is a clean product of the current pipeline. WCP and garvanbay are each ~20 sessions of hand-patches, surgical D1 stamps, grafts and partial regens layered onto seeds built by older code. Three consequences, all of which bit this session:

  1. We cannot answer "what does the pipeline produce today?" — so capability gaps get ranked from anecdote ("the video, the icons, the overlays… and many more") rather than from measurement.
  2. The scorecard measures a hybrid of pipeline output + accumulated hand-fixes, so its verdicts (garvanbay SHIP 92.2, WCP MANUAL-POLISH 79.0) do not describe the pipeline.
  3. Hand-written values masquerade as ground truth. A .tmp/*.sql patch put a WhatsApp button in WCP's CTA slot and a phantom social rail on the deployed site; the wiki then recorded those as "live (faithful)" and a whole slice was scoped against the inverted premise. See known-patterns "A hand-written D1 patch is a FABRICATION with a production address".

A clean rebuild answers (1), fixes (2), and makes (3) structurally impossible.

The deliverable is the gap report. The rebuilt sites are how we get it.

Decisions taken (Cathal, 2026-07-13)

  • Rebuild first, instrumented — not "fix the known gaps then rebuild". We stop ranking anecdotes.
  • Assets: reconcile + report only — do NOT bulk-migrate the 45 WCP videos. Without segmentation (which video sits in which section) they would not render; that is motion without progress.

Preconditions — DONE (2026-07-13 pt3, commit 555968d)

  • garvanbay is on the Wix API. wix_site_id: a8055256-… committed to apps/cms/sites/garvanbay.jsonc, verified against the API (41 files, 0 videos) — not taken on trust.
  • The media-pool instrument is fixed. searchFiles returns 0 for garvanbay (absent from Wix's media search index) — it would have certified the site as having zero asset debt. The pool is now mediaPool(), the union of search ∪ root ∪ every folder, warning loudly when the two enumerations disagree. This was the instrument the whole report depends on. See ADR-0010's second amendment.

In scope

  1. Safety. Dump both live D1s (ec_pages + ec_globals) to timestamped files; take Time Travel bookmarks. They hold the only copy of anything operator-authored (WCP's native block), and we are about to clear them.
  2. Baked-seed identity assertion. emit-seed.mjs per SITE, plus a build-time assertion that .emdash/seed.json#meta.name === SITE which refuses to build otherwise. Precondition: we are redeploying both workers, and today the WCP worker carries garvanbay's identity (see known-issues).
  3. Full rebuild from inception, both sites — fresh crawl → capture → theme → seed-from-build → image migration. Nothing hand-written survives.
  4. Clear + reseed both live D1s; redeploy both workers. Both sites become pure products of the current pipeline.
  5. Live-asset harvest (the new instrument). Playwright over the live pages, harvesting every media id referenced in the rendered DOM (the videos are JS-mounted, so the static crawl cannot see them — that is the whole bug). Join against the Wix pool to split at-source-never-pulled into:
  6. "used on live, our scrape missed it" → real capture debt (the 45 videos)
  7. "unused library cruft" → expected (old uploads, alternates)

This is the separation known-issues has been asking for since the 298 number appeared. 6. One ranked gap report per site — sections dropped, assets missed, props withheld, scorecard before/after.

Out of scope (these become the ranked backlog)

Fixing anything the report finds: the 45 videos, the fcr.usp-bar icons (icon: ""), the "Why choose us / VAT Registered" overlays, hero text layout, the floating-CTA orientation. Also out: the per-site static lane, the Wix API text plane (ADR-0010 slice D), gate calibration (Slice 6).

Done criteria

  • (a) Both workers serve a seed with zero hand-patches; the .tmp/*.sql lineage is gone.
  • (b) WCP's admin says WCP, and the build refuses to bake the wrong site's identity.
  • (c) garvanbay's asset reconciliation runs (it now can).
  • (d) at-source-never-pulled is split into two counted buckets, per site.
  • (e) A ranked gap report per site, with scorecard verdicts before/after.
  • (f) Both sites HTTP 200 — and any regression against today's baseline is reported, not hidden.

Risks (named up front)

  • garvanbay is SHIP (92.2) today and a clean rebuild may regress it. known-patterns is explicit: "Re-seeding a rebuilt CMS seed needs the AUTHORITATIVE basename→R2-URL map from object metadata — a fresh matcher rebuild loses image resolution, and the R2 key is a lossy hash you can't recompute." The rebuild must re-use that map. If garvanbay still comes back worse, that is a finding, not a failure — but the D1 dump is the restore path.
  • Brand nondeterminism on re-crawl. A fresh crawl can flip WCP's brand (source: logo → surface, confidence high → medium) and silently regress the palette (known-issues). Diff the derived theme, not just _raw, and stop if it moves.
  • originalNameOf returns a folder name (Misc) for 19 of WCP's 46 videos — a hard precondition of any video migration (which is out of scope here, but the report must not launder the bad names).

Baseline to measure against (as of 2026-07-13 pt3)

garvanbay WCP
scorecard S 88.9 / T 100 / G_det 90.8 → 92.2 SHIP S 70.8 / T 100 / G_det 74.3 → 79.0 MANUAL-POLISH
Wix pool 41 files, 0 videos 393 files, 46 videos (1 migrated)
loud-fail placeholders 0 0