Investigate Codex usage across interrupted, replayed and resumed turns #1946

Verify Codex usage accounting independently of context occupancy and subscription quota display.

A pure-parser probe shows codex.extractUsage returns null for a token_usage_record with payload.usage while a turn.completed direct usage control parses correctly. Establish whether that record shape reaches the actual app-server/live or transcript-import path before claiming a production undercount. Existing token_count total snapshots are already treated differently from request deltas.

Acceptance criteria:

• Exercise interrupted requests without following token_count, subsequent cumulative snapshots, replay of a sequence's first request, genuine resets, forks and restart/resume baselines.
• Count each measured request once; preserve original timestamp, provider, observed model and any known tier. Unknown values remain unknown.
• Keep cache/reasoning subsets and context-window occupancy separate from billable turn/session totals.
• Audit compressed .jsonl.zst discovery in session import, but do not claim the current application ledger aggregates every external transcript. Coordinate import work with Investigate: Rewound Codex history import #1906/Add projects from Claude Code / Codex / Cursor history #1858.
• Make skipped/inconsistent-record coverage visible if a transcript accounting path is introduced; deduplicate sources by identity and bound rescans/cache invalidation.
• Capture focused failures against the baseline before changing runtime accounting.

Coordinate with #1928 for per-model ledger storage and #1904 for limits/recovery. Tracked by #1926. Tests should use isolated fixtures without paid turns or real account data.