[ChristianPavilonis]

Status

Draft recovery candidate. Do not merge or close the issue yet.

The cacheable unified candidate is pushed at dec05bdc9c461a8e43c19d5f51be68e520ebb178. It improves the reproduced cold/warm lab regression, but acceptance of synchronous loading and the changed critical-script outage boundary is still pending. Dedicated real-GPT inventory is missing. The PR remains draft, with no review requests or issue-closing statement.

The earlier full CWV contract still requires deferred loading and rejects this candidate with loading must exercise its intended production loading mode. It has not been rewritten to make a diagnostic pass count as full acceptance. Framework/live consumer migration and the final acceptance matrix are deferred until the architecture decision. Hosted checks may therefore fail the unmigrated contract; full hosted success is not claimed.

Changes

• Preserve publisher-registered initial delivery, explicit payload admission, native-slot cancellation/destruction, targeting retirement and navigation generations. This is opt-in publisher cooperation, not transparent duplicate prevention for unchanged pages.
• Preserve explicit responsive slot-space reservation and its stable publisher-wrapper requirement. No format-derived guessed height or general zero-CLS promise.
• Use the existing versioned synchronous unified runtime for the supported GPT profile instead of embedding both raw and compiled early initializers in every HTML response.
• Keep configuration before the runtime, full GPT APIs before subsequent publisher scripts, actual CSP/diagnostics/trace handling, attribution/APS attributes and existing separately deferred integration order.
• Retain provisional initialization for other browser-module profiles unless actual enforced CSP requires omission. Keep implementation/handover code rather than expanding this into source or executable-size cleanup.
• Add two explicit test-only diagnostic scopes. Neither can satisfy the full acceptance verifier. No production loading flag, retry, codec or dependency is added.

Matched diagnostic measurements

Eighty measurements use pinned original main 51730bdaf841e4d7bb96c2258e5e98169bdff9a7, four alternating arms, two viewports and five samples per group. CPU is 4x, latency 100 ms, download 1,600 kbit/s and upload 750 kbit/s, with zero artificial runtime hold. The transport is uncompressed fictional loopback HTTP, not real GPT/GAM or field CWV.

Baseline to combined medians:

Condition	Viewport	FCP/LCP, ms	First visible ad estimate, ms
Cold	390x844	948 to 824	1,345.7 to 1,231.9
Cold	1280x1000	956 to 828	1,350.0 to 1,208.7
Warm	390x844	348 to 168	564.8 to 395.1
Warm	1280x1000	352 to 152	563.0 to 382.2

The candidate document is 2,104 bytes instead of the previous candidate's 128,941 bytes. Warm runs demonstrate real HTTP runtime cache hits and zero runtime transfer bytes. Each measurement checks one native fixture request, candidate targeting and a visible creative. Controlled CLS remains zero with the configured reservation.

All 14 production JS modules are byte-identical. Re-exporting the committed source proves all 11 servable files per arm and production source fingerprints match the measured artifacts. The screen retains its original tooling hashes; the later outage extension is separately executed, not relabeled as another timing run. Production compressed-delivery magnitude remains unverified.

Focused outage proof

Four separate browser cases cover normal and old-hash/self CSP success and critical-unified HTTP 404 under each policy. Success has working APIs, a single queue callback/request and visible creative. Failure has no TS initialization/shim/renderer APIs, no queue drain, one 404 without retry, zero native requests/delivery, completed parsing and only the declared core-dependent publisher error. CSP violations remain zero.

This differs from the previous inline/deferred design, where a failed deferred GPT asset left early rendering available. The diagnostic explicitly records changedOutageBehaviorApproved=false and fullAcceptancePassed=false. Observation is not acceptance of the new boundary.

Checks

Fresh checks pass:

• 3,376 native core tests, 20 native integrations and 30 doctests.
• 60 Axum tests and 28 cross-adapter parity tests.
• All eight target-specific Clippy aliases, adapter test aliases, native digest lint and native/actual Axum/Fastly EOF-auction checks. All 23 selected validation-script gates pass, including build, stamp and Rust formatting.
• 80 selected Node tests, zero skips, plus 16 compiled-IIFE startup tests.
• Wasm clock lint, Node pins, changed-domain lint, scoped ESLint/formatting and docs build.
• Two independent read-only cross-reviews, with no additional correctness findings.

Existing ignored tests remain ignored. Previous broad JS results are reused for unchanged production JS inputs, not reported as a fresh full suite. Standalone TypeScript's previously recorded inherited failures remain disclosed. The new candidate's complete framework and real-GPT acceptance have not run. The existing full CWV correctness scope exits 1 with zero completed samples because its deferral expectation is intentionally unmigrated.

Remaining decisions and proof

1. Accept synchronous loading with documented blocking dependencies, as issue Client CWV: defer tsjs bundle, CLS min-height injection, disableInitialLoad #854 permits, or retain the stricter parser-progress requirement and revise the design.
2. Accept the observed sole-critical-script outage boundary, or require the previous inline fallback guarantee and revise the design.
3. Once the contract is settled, migrate the affected full controls/framework/live consumers and run the final agreed matrix. Add deployment-relevant encoded delivery only where needed for those claims.
4. Supply owner-authorized dedicated GAM inventory, known TS/ordinary creative IDs and visible markers. Run genuine original-main failure reproduction and candidate real-GPT delivery. Missing inventory remains blocked, never skipped success.

No deployment, publisher activation or Google traffic is authorized by these diagnostic runs. Keep this PR draft even if later checks pass unless explicitly instructed otherwise.

Evidence

• Closure contract, historical regression and current recovery checkpoint
• Measurements and fingerprints. The recovery20261103 addition retains all 80 new rows and four outage observations; the original 200 rows remain unchanged.
• Diagnostic scope and commands
• Real-GPT prerequisites
• Related issue: Client CWV: defer tsjs bundle, CLS min-height injection, disableInitialLoad #854

[ChristianPavilonis]

Pushed 24a438e and updated the description and validation records. The matched tests expose a warm-cache regression: throttled mobile FCP/LCP rises from 348 to 756 ms and the visible-ad estimate from 550 to 1,008 ms, despite genuine runtime cache hits. Local functional checks pass, but that performance issue and missing dedicated real-GPT/GAM inventory prevent closure. This stays in draft without an issue-closing keyword or review requests until those blockers are resolved. The validation record and all 200 measurement rows are linked in the description.