Codex: dedupe fork-replayed history by content, not position - #37

Fork/spawn child rollouts replay the parent session's history with
every copied line's timestamp rewritten to the fork instant, so the
positional dedup key (session, timestamp, line index) counted replayed
token_count events again — inflating token totals, LIMITS samples, and
MODES effort mix (issue #36).

- Usage events: keyed on (session, last_token_usage vector,
  total_token_usage vector). The cumulative is non-decreasing between
  compactions, so the pair identifies one turn's consumption across
  every file it appears in; re-emissions with an unchanged cumulative
  collapse too. Logs without the cumulative fall back to the
  positional key.
- Rate-limit samples: same usage fingerprint plus the snapshot's own
  fields (used_percent, window_minutes, resets_at) so a verbatim
  replay collapses while a genuine re-notification survives.
- Effort events: keyed by turn_context.payload.turn_id, which replays
  copy verbatim; positional fallback for logs predating turn_id.
- merge_into: keyed duplicates keep the earliest observed timestamp —
  scan order (archived_sessions sorts before sessions wholesale) can't
  let a replay shift an event's day attribution. The five dedup loops
  fold into a shared dedupe_into helper.
- CACHE_VERSION 9 -> 10: cached FileEvents carry the old positional
  keys, so parsing semantics changed even though the layout did not.

Known remainder (follow-up): session touches and task durations from
replayed lines are still counted.

Co-Authored-By: Claude Fable 5 <[email protected]>

Address second-model review of the GH-36 fix:

- Gate the semantic-key path on a fingerprintable usage object
  (total_tokens parses as an integer) — total_token_usage null/{}
  would fingerprint as all-zero and falsely collide distinct turns.
  Same gate on the rate-limit key's last_token_usage.
- Clamp used_percent before keying so an out-of-range raw value can't
  mint a second key for the same stored sample value.
- Soften the CHANGELOG attribution claim: earliest-timestamp-wins holds
  when both original and replay are scanned; an out-of-window original
  leaves the replay's fork-instant timestamp (known remainder).
- Pin the review-raised invariants in tests: malformed-cumulative
  fallback, rate-limit re-notification survival, cross-turn re-emission
  collapse (why turn_id is not in the usage key), same-timestamp
  distinct turn_ids, cache_write discrimination, and Claude fragment
  merge (larger volume + earliest timestamp + merged metadata).

Co-Authored-By: Claude Fable 5 <[email protected]>