test(ai): repair exact dev evidence and timeout contract

[Yeachan-Heo]

Exact Dev CI repair

Repairs the two primary @gajae-code/ai failures from Dev CI run 30414283071 at exact dev head 0b1096861dea43ff18a4b1f8dcb0c4a05fd5780a. The evidence-producer and aggregate failures were downstream fail-closed cascades.

1. Canonical Anthropic cache evidence

artifacts/architecture-2383-eval.json carried source pins for blob a0acb862…, while exact dev's committed packages/ai/src/providers/anthropic.ts is blob 8dace169…. The mismatch already fails on pre-#3272 parent 1b5e2ad5; #3272's first-parent diff touches only Azure provider/test/changelog files.

Regenerated through the repository-supported path:

WRITE_ARCHITECTURE_2383_EVAL=1 bun test packages/ai/test/anthropic-cache-eval.integration.test.ts

The resulting official pins are:

• blob OID: 8dace16954264312b4bb99e78236f868ea6c1b8d
• SHA-256: 7f521d452360fd79b07f6d8f0ba7ef75e531d2289d444fce37686b21a81577e9
• input fixture SHA-256 remains b1ca8d3183ca168140774f848ed414252178f7c5788d61243069862ae59c129b

A non-writing rerun passes and an independent digest/blob check matches the committed provider bytes.

2. Deterministic Codex websocket first-event contract

The stale regression set removed PI_CODEX_WEBSOCKET_FIRST_EVENT_TIMEOUT_MS; commit dd27b1b2 intentionally removed that transport-specific variable and made first-event timeout a typed provider error for the outer retry/fallback layer. Consequently the test waited until Bun's unrelated 5000 ms test deadline.

Bounded reproduction before repair:

• pre-fix(ai): resolve the Azure API key fallback from trusted env #3272 parent 1b5e2ad5: 10/10 timed out at ~5001 ms
• exact current dev: 22/22 timed out at ~5001 ms before the interrupted bounded run ended
• fix(ai): resolve the Azure API key fallback from trusted env #3272 has no diff overlap with the implementation or regression file

The test now supplies streamFirstEventTimeoutMs: 10 directly and pins the current safety contract: one websocket send, typed stream_first_event_timeout, no provider-internal SSE replay, and unchanged websocket session state. Post-repair stress: 25/25 pass in under one second. No timeout inflation and no production-code change.

Verification on rebased head 70d4fdfe44cb02c0236b3dfb8387447bc0f8ddef

• Focused five-file repair net: 73 pass / 0 fail
• Repeated deterministic websocket regression: 25/25 pass
• Full @gajae-code/ai suite under CI-equivalent neutral trusted env: 2005 pass / 337 skip / 0 fail across 231 files
• bun run --cwd packages/ai check: pass (Biome + TypeScript)
• bun run build: pass for all build-bearing workspaces, including native and compiled coding-agent binary
• git diff --check: pass

A plain local full-suite run inherited this workstation's configured trusted provider endpoints and produced eight existing configuration-sensitive trust-test failures; the neutral trusted-env run above matches CI isolation and is the authoritative full-suite result.

No product source, secrets, workflow gates, or unrelated subsystem files are changed.

[Yeachan-Heo]

Independent exact-head verdict: APPROVE — PR #3457 head beea55994c93461f82f7c3e80a74f4e336205d76, parent/current dev 0d13c7486cc7e9e47eeddc1a67ab9156a443384a.

The final independent architect/red-team pass is clear. It specifically re-tested the prior high-severity review concern by static adversarial analysis: the timed-out Codex websocket is closed only when its waiter expired with an empty queue; the next outer retry therefore creates a fresh socket, while late frames remain confined to the discarded connection. The regression is non-tautological: it emits a stale response from request 1 before the delayed fresh response from request 2, and would fail without the production close. It also pins typed stream_first_event_timeout, no provider-internal SSE replay, no websocket disablement/fallback-count increase, a fresh second socket, and fresh-only output.

Evidence integrity is exact: artifacts/architecture-2383-eval.json records Anthropic provider blob 8dace16954264312b4bb99e78236f868ea6c1b8d and SHA-256 7f521d452360fd79b07f6d8f0ba7ef75e531d2289d444fce37686b21a81577e9, matching the committed source and the repository-supported WRITE_ARCHITECTURE_2383_EVAL=1 regeneration path.

Terminal gates:

• Exact-head CI: 23 SUCCESS, 6 SKIPPED, 0 pending/failing; mergeability clean.
• Exact current-dev CI run 30417195980: terminal SUCCESS at 0d13c7486.
• Local repeated stale-frame regression: 25/25 pass.
• Focused repair net: 73/73 pass.
• Full @gajae-code/ai suite on the exact rebased head under CI-equivalent neutral trusted env: 2009 pass / 337 skip / 0 fail across 232 files.
• packages/ai Biome + TypeScript check: pass.
• Full workspace build, including native and compiled coding-agent binary: pass.

No unresolved finding remains. Exact head is safe to merge.

— GJC security/trust lane, independent architect/red-team review, 2026-07-29