fix(combo): auto-resume pinned native Codex turns

[mdigitalbh81]

Summary

Follow-up to #12240.

Native Codex turns remain pinned to the same provider/model while eligible sibling connections for that model still exist.

If every target for the pinned provider/model becomes unusable for a model-scoped reason, such as quota exhaustion or model lockout, this change safely advances the logical turn generation and resumes the same Codex turn on another healthy model from the combo.

The resumed model is then pinned for the new logical generation, so subsequent requests for the same thread_id + turn_id continue consistently on that model.

Safety constraints

• Pin scope remains thread_id + turn_id; it is not global.
• Sibling connections for the same pinned provider/model are preferred before cross-model resume.
• Auto-resume is limited to model-scoped unavailability.
• Provider-wide circuit breaker/cooldown conditions do not trigger cross-model resume.
• Pending tool calls or opaque/provider-specific continuation state reject auto-resume.
• Partial-stream failure does not dispatch another model mid-stream.
• At most one auto-resume is allowed per logical turn.
• Failed auto-resume dispatch does not advance generation or cascade to a third model.

Implementation

• Adds generation-aware Native Codex turn records and pinning.
• Adds conservative eligibility checks for safe cross-model continuation.
• Adapts the auto-resume state into the current extracted combo attempt-loop architecture.
• Advances the generation and re-pins only after a successful alternate-model response.

Validation

Focused auto-resume suite:

• tests/unit/native-codex-auto-resume.test.ts: 15/15 passed, run twice

Regression coverage:

• existing Native Codex turn-pin regression suites: 15/15 passed
• npm run typecheck:core: passed
• ESLint on all 5 changed files: passed
• git diff --check upstream/release/v3.8.51...HEAD: passed

Live validation was also performed on ARM64 with a real Antigravity Opus 429/quota exhaustion: the pinned turn became eligible for auto-resume, routed to Gemini, completed with HTTP 200, and subsequent requests for the same turn remained pinned to Gemini.

[diegosouzapw]

Thorough follow-up to #12240 — the safety constraints you list in the PR body aren't
just words, each one has its own dedicated test (circuit-breaker/cooldown non-trigger,
pending-tool-call block, opaque-continuation-state block, sibling-connection
preference, the 1-per-turn cap, dispatch-failure non-cascade), and the "5-Phase
Production Scenario" test is a great piece of documentation on its own. Ran
native-codex-auto-resume.test.ts at your head: 15/15 pass; eslint on the touched
files is clean. One thing before merge: please add a changelog fragment under
changelog.d/fixes/. Also flagging for the maintainer queue (not something you need to
act on): a few other open PRs in this batch touch executeTargetAttempt.ts too, in a
different hunk than yours — should merge clean either order, but worth a merge-tree
check at merge time.

[diegosouzapw]

Almost there on this one (⭐4 in review). Two pre-merge items remain, and the first one needs you
because this fork has "Allow edits by maintainers" unchecked — our fix-sweep normally pushes
small items like this directly into the contributor's branch (authored by you, maintainer as
co-author):

1. A changelog fragment under changelog.d/fixes/ (one - bullet describing the auto-resume of
   pinned native Codex turns, ending with (#13180)).
2. At merge time we'll re-run git merge-tree against the current release/v3.8.51 tip and
   against the other executeTargetAttempt.ts PRs that landed meanwhile (fix(combo): retry pre-content streaming failures #13630, [defer] fix(combo): hop Astra quality-empty 200 and Gemini 3.8 empty-content to family siblings #13603, fix(resilience): clear the combo LKGP pin only when it names the failed target #12235)
   — that one is on us, nothing for you to do.

Either tick "Allow edits by maintainers" and say so here, or push the fragment yourself and we
take it from there.

[diegosouzapw]

Thank you, @mdigitalbh81 — this landed. Your fork does not accept maintainer pushes, so the work was re-landed as #14162 with your commits carried intact (authorship preserved) and squash-merged into release/v3.8.51 as cd4c6f69. On top of your commits the maintainer side added the reconciliation with #13564 (which had changed the same pin-release path meanwhile), a discriminated-union narrowing fix in combo.ts, and a split of the 1540-line test into two files under the cap — all described in the #14162 merge body. Closing this one only because the same change is already on the release branch under your name.

[diegosouzapw]

Your native Codex auto-resume fix is on release/v3.8.51: it landed through #14162 (a maintainer re-land, since the fork did not accept maintainer pushes), squash cd4c6f69. That squash was authored under the maintainer's name, which was not the intent — commit 06f1df9d (#14362) now carries Co-authored-by: Felipe Reis so the credit is recorded in history, the contributors graph and the release notes. Closing as landed. Thanks for the fix, Felipe.