Two-issue gateway-recovery instrumentation registration

Registered before these issue solver and review calls. This small, separate instrumentation pilot checks the revised gateway policy and two newly selected public issues; it is not the requested 30–50-task cohort. Selection followed the previous batches and observed route failures. The manifest has SHA-256 c5bbca09a6c7005b448d8e2ee8e8c9f71f2d47e723d57604dbb53c6330290a84 and freezes issue and context hashes, checks, code, models and randomized paired order. Both issues were open with complete comment snapshots when captured and were rechecked before registration.

Issue	Frozen issue checks	Positive reference and controls
Prettier #16126	The reported folded YAML and one renamed/reworded variant preserve their parsed value under the stated options and remain idempotent.	The source-only patch from PR #20292, head 57a33a438d8eb8bd1e4b4c88dd3922421ec924c2, passes the checks. Uncompressed patch SHA-256 6690c0c7af0f3b13c8886a2abb3047dc40f770124cf0b8dd02203595c9dfef4b. Three small YAML controls pass on both baseline and reference.
Prettier #17487	The reported starred JSX comment and one variant keep the asterisks aligned, retain comment text and normalized JavaScript syntax, and remain idempotent.	PR #17494, head 95fb9dfb12c770fa35066654613300e62ca44f4e, did not apply to the pinned source. A manual source-only port passes the checks; its uncompressed SHA-256 is fb9d37bb3884e1085893fcb51df4024855371ffe73859fc3063ae486843a7f4a. Three small JS/JSX controls pass on both baseline and port.

Both issue checks failed on pinned Prettier bed4edbb1cf2a9db4004215b0ce5707cdb02b5d9 and passed on their references before registration. The public test exercises the original reported example; the acceptance check also uses a closely related variant. The controls are partial: they do not constitute the full Prettier test or lint suites or maintainer approval. The YAML reference changes a different function from the prior #20293 reference but touches the same source file; issue independence is not established. The JSX reference is a local port, not an upstream merge. Public PRs and comments may disclose solutions or be represented in model training.

The solver is gpt-6-astra; the blinded AI reviewers are deepseek-flash and claude-haiku-4-5, and the AI adjudicator is qwen3.8-flash. Each name is accessed through the same proxy; provenance and independence are not verified. A separate small synthetic Qwen review-format probe passed before registration, which does not establish reliability for these tasks. The same VM, bounded search/read/edit tools, 12-turn solver limit, explicit review rubric and per-stage stop rule as the earlier batches apply. The two retrieved contexts use the same maximum limits but have different content and actual token counts; the issue histories are short and pre-seeded, not long production sessions or autonomous memory management.

For this new pilot only, an explicit HTTP 503 may be retried at most twice, with 2- and 8-second fallback delays, respecting a Retry-After up to 30 seconds. A longer Retry-After stops the operation. Every attempt is recorded separately; a timeout or other ambiguous failure is not automatically retried. After exhausted retries or another provider error or missing usage, the current stage stops launching requests. The earlier no-retry registrations and their results are unchanged. The estimated model-call target is below $1 for this small pilot, not a hard cap; failed and unreported charges remain unknown. A successful retry cannot prove the upstream cause of 503 has been fixed.