fix(codex): fence overlapping compaction attempts

[mulfyx]

Summary

Codex compaction kept one registry entry per Claude session, while asynchronous compaction producers and terminal handlers identified that entry only by session and model.

On base 529e1d6, a compaction request awaits upstream work before storing its result (request path). Meanwhile, store_compaction replaces the session slot unconditionally, and activation/abort do not identify the originating attempt (registry and store, activation and abort).

With overlapping compactions A and B for the same session, a late A completion could therefore:

• overwrite B's newer native history;
• anchor B's native history with A's portable summary;
• clear B's state after a late A failure.

This change:

• allocates a monotonic CompactionAttempt before awaiting upstream compaction;
• publishes a Preparing generation immediately;
• permits store, activation, and abort only for the current generation;
• carries the generation through replay and all buffered/live completion and cleanup paths.

The scope is same-session stale-completion/lost-update protection, not a cross-session isolation issue.

Consistency precedents

These are consistency precedents, not claims that the Responses compaction API mandates generation IDs:

• This repository's continuation state already allocates a turn ID before asynchronous work and rejects stale publish/abort operations (allocation, guards).
• OpenAI documents compacted encrypted_content as an opaque item that must be carried into subsequent Responses requests (compaction guide). Keeping it paired with the portable summary from the same attempt is therefore an application-level consistency requirement.
• OpenAI Codex uses expectedTurnId as an optimistic concurrency precondition (schema) and validates it atomically under the active-turn lock (implementation).
• RFC 9110 describes conditional writes as protection against parallel lost updates (If-Match section 13.1.1).

Validation

• The stale-activation regression test fails on the unfenced implementation and passes with this change.
• Deterministic tests cover stale store, activation, abort, replay abort, invalid stale summary, and replay checks while a newer generation is still preparing.
• cargo test --all-targets: 828 tests passed.
• cargo clippy --all-targets -- -D warnings: passed.
• A controlled runtime A/B race delayed A until B completed; the next turn replayed the newer native-B history.