[Bug]: Codex usage parser does not reconcile per-event and cumulative counters #5800

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Create a Codex rollout containing a turn_context followed by a valid token_count event whose total_token_usage advances by its last_token_usage.
2. Add another token_count event where the cumulative total_token_usage is unchanged but last_token_usage differs, as can happen when Codex recomputes current-context usage or re-emits token state alongside rate-limit updates.
3. Alternatively, make total_token_usage advance by a value that does not equal last_token_usage.
4. Scan the rollout from the Usage page.

The Codex protocol maintains total_token_usage by adding each real last_token_usage element-wise. An unchanged cumulative total therefore proves that a later event did not add billable usage, and a different cumulative delta proves that the two counters are inconsistent.

Expected behavior

The usage parser should use total_token_usage as an integrity signal. It should count a legitimate per-request last_token_usage once, suppress events whose cumulative total did not advance, and report inconsistent counter sequences as malformed rather than guessing which counter is correct.

Reasoning tokens must remain a subset of output tokens and must not be added to the processed-token total again.

Actual behavior

parseCodexLine never reads total_token_usage. It trusts each nonzero last_token_usage whose JSON signature differs from the immediately previous one.

This means a stale or recomputed last_token_usage can be counted even when the cumulative total did not advance, and a malformed per-event counter can be accepted even when it cannot reconcile with the cumulative counter. The Usage feature's original PR stated that these fields had been reconciled before last_token_usage was trusted, but that invariant is not implemented in the parser.

This is separate from copied parent history in forked rollout files (#5758): the issue here occurs within one rollout's counter sequence.

Impact

Major degradation or frequent failure

Version or commit

main @ 963ebf5

Environment

Platform independent; Codex rollout transcript parsing in apps/server

Logs or stack traces

No error is logged. Inconsistent token counters are accepted silently.

Screenshots, recordings, or supporting files

No response

Workaround

Use a usage parser that validates each per-event counter against the cumulative Codex counter.