fix(proxy): retry accepted capacity bridge failures

[Komzpa]

Summary

• retry accepted-but-output-free HTTP bridge overload, selected-model-capacity, and abrupt-close failures before model output is visible
• retry direct non-streaming Responses API overload responses and initial upstream stream EOFs through the existing bounded retry path
• preserve owner-pinned quota handling while keeping local bridge/websocket close failures account-neutral
• bound accepted retries by downstream attachment, request deadline, and visible-output state; retain the pending close failure for the reader when replay cannot proceed
• update OpenSpec coverage and CI label-helper tests for retry-server-is-overloaded

Root cause

The bridge handled failures before response.created, but later accepted responses could still fail with overload/capacity or an abrupt close before any model output became visible. Those failures were surfaced as terminal even though the narrow output-free request was safe to replay. The direct non-streaming and initial-stream-EOF paths similarly returned before entering their bounded transient retry path.

The retry path now preserves downstream continuity and refuses replay after downstream detach, deadline expiry, visible output, or an unsafe continuation. Local transport and reader closes remain account-neutral; real upstream account errors retain their existing health penalties.

Validation

• Head: 4ee8538421eaa6a46cfa199055b435dd019deb45
• uv run pytest tests/unit/test_proxy_http_bridge.py tests/unit/test_proxy_utils.py -q -> 1239 passed
• uv run pytest tests/integration/test_http_responses_bridge.py tests/integration/test_proxy_api_extended.py tests/integration/test_proxy_transient_retry.py -q -> 183 passed
• uv run ruff check app/modules/proxy tests/unit/test_proxy_http_bridge.py tests/unit/test_proxy_utils.py tests/integration/test_http_responses_bridge.py tests/integration/test_proxy_api_extended.py tests/integration/test_proxy_transient_retry.py -> passed
• uv run ty check app/modules/proxy -> passed
• openspec validate retry-server-is-overloaded --strict -> passed
• git diff --check origin/main...HEAD -> passed

[Komzpa]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: aabc8da834

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Komzpa]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8062265485

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Komzpa]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Already looking forward to the next diff.

Reviewed commit: 8043fdc51e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Komzpa]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Delightful!

Reviewed commit: 7563c322f7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Komzpa]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0148a2f96a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Komzpa]

@codex review

[Soju06]

Thanks for the 08-14 rebase — conflicts are cleared, all review threads are resolved, and the eventless-deadline reset landed in _prepare_http_bridge_terminal_capacity_replay. Remaining blockers before this can go to a fresh review pass:

1. CI Required is red on head 7202460a: ty fails in tests/unit/test_key_fingerprint.py:34 (Expected AsyncSession, found _FakeSession).
2. Out-of-scope changes folded in by the rebase: app/core/config/key_fingerprint.py (sqlite busy-retry) + its test, and the websocket/mixin.py reader-await/cancellation hunks carried over from the now-closed fix(proxy): retry websocket EOF before visible output #1388. Please split these into their own PR(s) — the key_fingerprint change is also what breaks CI here.
3. Half of the 07-30 item 1 is still open: the reset is in the prepare helper, but latency_response_created_ms is not saved/restored in the finally block of _retry_http_bridge_terminal_capacity_request (a failed replay now leaves it None while the rest of the state is rolled back), and the requested regression test (replayed socket accepts the send and goes silent -> eventless deadline fires) is missing.
4. Fixes #N line is still absent from the PR body.
5. Once 1-4 are done, trigger a fresh automated review on the rebased head — the last review predates the rebase, and note the repo is moving review to CodeRabbit (chore(github): retire codex review label gate in favor of CodeRabbit #1763).

Also flagging: main's #1713 archived retry-server-is-overloaded, so this PR now edits docs under openspec/changes/archive/2026-08-13-retry-server-is-overloaded/ — maintainer should confirm that editing the archived change dir is the intended reconciliation.

[Soju06]

Rebase nudge: this was one review pass from mergeable on 08-16; now needs the rebase plus the ty fix. Still wanted.