fix: recover Codex QA follow-ups without rollout

[numnx]

What changed

• Detect Codex thread/resume failures that specifically report no rollout found for thread id from structured provider errors or stderr.
• Let Task QA coding follow-ups preserve their existing task workspace and retry once as a fresh Codex conversation.
• Keep Codex fallback opt-in so strict restart/planning continuations and unrelated conversation flows remain fail-closed.
• Preserve existing Claude Code and OpenCode fallback behavior.
• Update canonical and published QA/orchestration documentation.

Why

Task QA can request a coding follow-up after a task completes. The handoff correctly retained the coding session and workspace, but a persisted Codex native thread ID could outlive its workspace-local rollout. Codex then rejected thread/resume, causing the Task Coding invocation to fail before applying the QA fixes.

The QA follow-up prompt is self-contained with the original subtask and concrete QA instructions, so retrying only the provider conversation while retaining the workspace safely recovers this missing-runtime-artifact case.

User impact

A single-task QA review that requests changes no longer fails solely because Codex cannot find the persisted rollout. The task stays in its QA handoff, keeps its code/worktree, and receives one bounded fresh-session retry. Unrelated Codex failures remain visible and are not retried by this path.

Validation

• pnpm run lint — passed
• pnpm run test:backend — passed (467 files; 6,163 passed, 12 skipped)
• pnpm run build — passed
• Focused regression suite — passed (4 files; 261 tests)
• git diff --check — passed

Risk and rollback

Risk is limited to the exact Codex missing-rollout error during explicitly opted-in continuations. The retry is bounded to one attempt and preserves the existing workspace. Roll back commit ee81177d8 to restore the prior fail-closed behavior.