v4.0 QC Sweep Review — 2026-05-14 (v1.4 · corrected-diagnosis cleanup · 7 perf decisions)

🛡 Shipped since the sweep (2026-05-15)

Canonical perf budget — reference/page-performance-budget.md. HARD ceilings (block publish): total ≤ 2,000 KB · script ≤ 1,000 KB · media ≤ 600 KB above-fold · perf ≥ 50 · LCP ≤ 4.0 s · CLS ≤ 0.25. Plus image/video/script/markup rules + the 6-step pre-publish checklist.
Pre-publish Lighthouse gate — tools/lighthouse_gate.py. Runs Lighthouse, asserts every threshold, exits 3 on HARD violation. Bypass requires LIGHTHOUSE_GATE_BYPASS_REASON + auto-logs to reference/deployment-log.md. Smoke-tested: correctly flagged UK brain-fog quiz with 6 HARD + 1 SOFT violations (LCP 15 s, 2.8 MB).
Per-page perf diagnosis — corrected. The first pass (PERF-DIAGNOSIS.md v1) guessed a "hero MP4" mechanism and got it wrong — there is no hero MP4. PERF-DIAGNOSIS-v2.md (ground-truth, Codex-QC'd) supersedes it: the video weight is an already-optimised bz-slide-video testimonial carousel, and the real problem is fleet-wide — 55 of 77 pages fail the HARD gate, dominated by LCP from third-party app JS. See the Corrected diagnosis section below.
Wired into the playbook — AGENTS.md MANDATORY pre-deploy gates list + ops-shopify-page-builder pre-flight check #14 + ops-quiz-deploy. Banned in memory at feedback_page_performance_budget.md.

Commit bf981a4 on feature/v4-migration · 6 files / 692 insertions

Outcome at a glance

88 v4.0 published registry rows → 77 unique handles after cross-market dupe collapse. Each page passed through three independent probes: Lighthouse (desktop perf/SEO/a11y/best-practices), Playwright mobile (console + render + screenshot), and raw HTTP (canonical/hreflang/JSON-LD/og/meta). Composite 0-100 per page. No FAILs.

Total pages QC'd

PASS (≥80)

WARN 70-79

WARN 60-69

73.7

Mean score

231

Cells written to Sheet

Systemic findings (theme-wide patterns)

What the sweep surfaced: the issues below are not page-specific; they recur across the v4.0 theme. Fixing the root once removes the finding from dozens of pages at a time.

100%

Console errors

77/77 pages have ≥1 JS error at runtime — theme-wide

65%

Accessibility < 90

50/77 pages — likely shared template a11y gap

52%

No JSON-LD structured data

40/77 pages — SEO gap; rich-result eligibility

43%

Performance < 50

33/77 pages — LCP / TBT pressure under 4G profile

25%

OG image / title missing

19/77 pages — social-share preview broken

12%

Slow load (>15s)

9/77 pages — tail latency outliers

99%

SEO score ≥ 90

76/77 pages — on-page SEO basics are solid

100%

HTTP 200

77/77 reachable — no orphan / 404 / 500 pages

✓ Banked decisions — logged 2026-05-15

All 8 decisions banked — submitted by Calum 2026-05-15 (see projects/briefs/qc-v4-sweep-2026-05-14/DECISIONS-LOG.md). Every card below is shown as BANKED ✓ — settled. The original slow_load → PROFILE_TOP3 verdict was superseded by the diagnosed slow_load_diagnosed card.

#	Decision	Verdict	Status
1	Console errors on 100% of pages	AUDIT_THEME_JS	BANKED
2	JSON-LD missing on 40 pages	TEMPLATE_PER_TYPE	BANKED
3	23 exam-season city LPs	REFACTOR_TEMPLATE_NOW	BANKED
4	Accessibility < 90 on 50 pages	AXE_AUDIT	BANKED
5	OG image/title missing on 19 pages	TEMPLATE_DEFAULT	BANKED
6	qc_result schema format	KEEP_HYBRID	BANKED
7	9 slow pages — diagnosed (ship which fixes)	SHIP_ALL_NOW ⚠ plan later invalidated — see Corrected diagnosis	BANKED
8	Lighthouse perf gate — adoption scope	RETROACTIVE_FULL	BANKED
—	9 pages slow load (original framing)	PROFILE_TOP3	SUPERSEDED

⚠ Corrected diagnosis — the perf fix plan changed

The original PERF-DIAGNOSIS guessed the mechanism and got it wrong. Ground-truth investigation of the live themes (2026-05-15) found:

No "hero MP4" exists. The video weight is an already-optimised bz-slide-video testimonial carousel (8 videos on exam pages, 34 on dopamine-guide) — click-only play, posters, preload="metadata", a prior Codex desktop-preload fix. The v1 "add preload + poster" fix was already done.
It's not 9 slow pages — it's 55 of 77 failing the HARD gate, dominated by LCP>4s fleet-wide incl. pages with no video.
Theme JS audit: the Brainzyme theme code is clean (scripts deferred). The cause is third-party app JS — ~516 KB unused JS + ~4.4 s JS execution per page from gtag, Judge.me, Reamaze, Meta Pixel, Hextom, Shopify Web Pixels Manager. App-injected — not fixable in theme code.

Live deploys were halted — the original SHIP_ALL_NOW authorised an invalidated plan. Codex 5.5 decision-QC confirmed the halt (RETHINK). Full detail: PERF-DIAGNOSIS-v2.md · THEME-JS-AUDIT.md · FINAL-PLAN.md (the last reviewed by Codex 5.5 + Gemini, both REFINE, all points folded in).

✓ Decisions — all 15 banked

All 15 cards banked ✓ — 8 QC decisions + D1–D7 performance decisions, all settled 2026-05-15. The performance workstream is now parked — Calum pivoted to the ads-go-live path. Remaining app-stack actions (Judge.me, Reamaze, Web Pixels audit, Consent Mode v2) are tracked as app-side tasks in pending-tasks-registry.md §v4.0 Performance. Cards render read-only below with the decided verdict.

🛠 Corrected fix plan — 3 workstreams

The v1 fix plan (Groups A–E) was invalidated. It was built on PERF-DIAGNOSIS.md v1's wrong "hero MP4" mechanism — the preload/poster fixes it proposed are already in place. The corrected plan is in PERF-DIAGNOSIS-v2.md + FINAL-PLAN.md (Codex 5.5 + Gemini reviewed). The performance workstream is now PARKED — Calum pivoted to the ads-go-live path; the live D1–D7 app-stack decisions above carry the real fixes.

Workstream	What it actually is	Risk	Status
WS1 — Carousel mobile video	`bz-slide-video` Embla carousel keeps `preload="metadata"` on mobile (8–34 videos/page). Already optimised on desktop (click-only play, posters, Codex-reviewed preload fix). Needs a real mobile-device test before any change — the component is shared across 55+ live pages mid-T1.	High care	PARKED
WS2 — Theme JS baseline	The likely dominant fleet-wide LCP driver: ~440–587 KB unused JS + ~4.4 s JS execution per page from third-party apps (gtag, Judge.me, Reamaze, Meta Pixel, Hextom, Web Pixels). App-injected — fixed app-side via the D1–D6 decisions above, not in theme code.	App-side	PARKED
WS3 — JSON-LD + CLS	40 pages emit zero JSON-LD (additive Liquid, low risk). `sections/bz-exam-city.liquid` `<img>` tags lack `width`/`height` — the CLS fix. Two small reviewed batches; the only genuinely low-risk work.	Low	PARKED

🏆 Top 10 — strongest pages

#	Mkt	Wave	Handle	Score	Perf/SEO/A11y/BP	Live URL

🚨 Bottom 10 — pages needing attention

#	Mkt	Wave	Handle	Score	Perf/SEO/A11y/BP	Findings	Live URL

All 77 pages — per-row review

Market Verdict Has finding

Mkt	Wave	Type	Handle	Score	P/S/A/B	Findings	Decision	Feedback

Methodology — how the score is calculated (composite 0-100)

30 pts — HTTP probe: status 200 (8) + canonical (5) + hreflang ≥3 (5) + JSON-LD ≥1 (6) + og:title+og:image (3) + meta title+desc (3).

40 pts — Lighthouse: perf×0.15 + seo×0.10 + a11y×0.10 + best-practices×0.05.

30 pts — Playwright mobile: nav 200 (10) + zero console errors (12, sliding) + load <8s (8) / <15s (4).

Verdict tiers: PASS ≥ 80 · WARN-A 70-79 · WARN-B 60-69 · FAIL < 60.

What's already been done as a result of this sweep

Pages Registry Sheet: 231 cells written (QC Result / QC Findings / Last QC) for 77 v4 rows by Handle match.
Go-Live dashboard: A/B link + LP-LIVE/DRAFT chip + QC score collapsed into a single two-tier widget per variant. Hover shows Lighthouse breakdown + findings + last-QC timestamp.
data.json refresh rule: every Sheet-write now MUST call tools/registry_pull_after_write.py in the same script — codified in AGENTS.md, ops-pages-registry SKILL, and memory/feedback_sheet_write_auto_pull.md.
qc_result schema v1.7: accepts numeric-suffix variants PASS|WARN|FAIL (NN/100) alongside the lowercase enum. Schema warnings on pull dropped 69 → 2.
GSD hook stubs: 4 missing hook scripts at C:/Users/PC/.claude/hooks/ stubbed to suppress the spurious MODULE_NOT_FOUND noise on every Write/Edit.