Fix Codex resume fork detection rejecting same-cwd rollout with prompt lineage but no shared item ids

[Juliusolsson05]

Problem

A resumed Codex pane can show transcript unavailable even while the conversation mostly continues rendering.

Bundle:

/Users/juliusolsson/.config/agent-code/debug-bundles/2026-05-17T12-04-00-107-b6c85d86

Error in state snapshot:

Codex resume: ignoring same-cwd rollout /Users/juliusolsson/.codex/sessions/2026/05/17/rollout-2026-05-17T13-45-54-019e35c1-cff4-79e1-aff8-e96a7d43b61b.jsonl — only 0/3 item ids shared with resumed file /Users/juliusolsson/.codex/sessions/2026/05/16/rollout-2026-05-16T19-51-18-019e31e9-fa43-7bc1-af2b-3c65c96d39cf.jsonl; treating it as an unrelated session.

Observed effect:

• state-snapshot.json has transcriptStatus: "error" and the above transcriptError.
• The UI still renders some of the conversation, but user prompts can be missing because the committed transcript channel is no longer confidently attached to the correct rollout file.
• The rejected candidate has the same cwd and contains the same user-visible conversation lineage, but the item-id overlap guard sees 0/3 shared opaque ids.

Why this happens

packages/codex-headless/src/CodexHeadless.ts accepts a newly-created same-cwd rollout during resume only when it shares enough opaque rollout ids with the original resumed file. That is too strict for resumed/reconstructed/compacted Codex rollouts: Codex can preserve the user-visible conversation while minting new response/item ids, so opaque-id overlap can be zero even for the correct successor file.

Likely fix

Extend resume fork detection beyond opaque item ids:

• keep cwd + freshness checks
• keep item-id overlap as the strongest signal
• add a secondary user-visible prompt lineage signal from rollout response_item user messages / event_msg:user_message
• accept a same-cwd candidate when recent prompt overlap is strong enough, even if opaque item-id overlap is zero
• keep diagnostics explicit so sibling-agent rejection remains debuggable

This probably belongs in codex-headless first, then an Agent Code submodule bump.