fix(server): bind Cursor retry replay to the admitted OAuth account on 429 failover

[luvs01]

Summary

• applyFailoverSnapshot commits through commitResolvedOAuthSelection, which can admit a DIFFERENT account than the proposed one when a newer manual selection wins the proposal race. The helper already re-stamps route.provider, sentOAuthSnapshot, and replayOAuthCredentialSnapshot from the committed snapshot — but the runTurn preflight 429 rotation in run-turn-execution.ts still passed the REJECTED proposal's { accountId, generation } to bindRouteReasoningReplayScope, binding Cursor's credential-scoped replay/checkpoint identity to an account that never served the retry.
• Return the admitted OAuthAccessSnapshot (or null) from applyFailoverSnapshot so the caller binds the replay scope to the same account the retry actually uses.
• Extend the adapter-event failover test to observe continuation-owner credential identity and assert the replay is scoped to the admitted manual account, not the rejected proposal.

Verification

• Head: 10cb5f6e93ae16170694292295e86c471b71afcf on dev 6467235a85df0f649a588c601532316d6ae0fce9 (merged 2026-09-18; clean ort merge, no conflicts).
• bun x tsc --noEmit — clean.
• bun run structure:check, bun run privacy:scan, bun scripts/file-size-ratchet.ts — passed.
• bun test tests/oauth/adapter-event-oauth-failover.test.ts — 12 pass / 0 fail. The new assertions fail against the unfixed source: the replay attempt's continuation-owner credential identity matched the rejected proposal, not the admitted account.
• bun test tests/oauth/generic-oauth-failover.test.ts tests/lib/reasoning-replay-scope-source.test.ts — 45 pass / 0 fail.
• Fork CI: diagnostic run 35291796244 (contributor fork Actions)

Summary by CodeRabbit

• Bug Fixes
  • Improved OAuth account failover when a manual account selection occurs during a pending request, ensuring the most recently selected account is honored.
  • Preserved conversation continuity by binding retries to the account and session state actually admitted after failover.
  • Added coverage for credential consistency across ongoing conversations and correct credential changes after switching active accounts.

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 9b103629-32bd-4dce-8de0-de80ee39e988

📥 Commits

Reviewing files that changed from the base of the PR and between f3cb520 and ddd630d.

📒 Files selected for processing (3)

• src/server/responses/request-transport.ts
• src/server/responses/run-turn-execution.ts
• tests/oauth/adapter-event-oauth-failover.test.ts

 ____________________________________________________________________
< Never go full rewrite. You don't buy that? Ask Netscape Navigator. >
 --------------------------------------------------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

✅ Deterministic PR hygiene checks passed.

[github-actions]

✅ READY

• all PR quality gates passed.

Hygiene

✅ Deterministic PR hygiene checks passed.

[lidge-jun]

Rebased this branch onto current dev (f3cb520400) and force-pushed. No change to the diff; the rebase is the whole fix.

test 2/4 was failing on the file-size ratchet, and the offender was tests/server/management-provider-validation.test.ts at 5529 lines against a cap of 5506 — a file this branch does not touch. It branched from a dev commit that carried that state, and dev cleared it afterwards in #5018 by moving the newest cases into sibling files. The file is 5498 on current dev, so the branch inherited a defect that no longer exists.

Verified after the rebase: no file in the baseline exceeds its cap in this tree. CI re-approved at the new head.

[lidge-jun]

macos 1/2 reported cancelled at 20m16s against the 20-minute platform-macos budget, and the aggregate correctly refused to call that a pass. Same contention I created by approving nine Cross-platform runs at once on an account that runs one at a time; the same shard finished inside budget today on dev, #4926, #4938, #4994, #5050 and #5042.

Re-running the failed job rather than widening the budget.