Count Codex token_usage_record rows and skip repeated token totals

[starkdmi]

Behaviour

Codex rollout parsing now matches CLI 0.147–0.155 usage lines.

• Repeated token_count snapshots no longer inflate totals. When total_token_usage is present and equals the previous total in the same file, that line contributes no usage. A second copy for another limit_id (for example premium) is the same case. The line still produces a quota observation, with no usage sample. A missing total_token_usage still uses last_token_usage. A smaller total (fork or resume reset) still counts as new usage.
• The phantom token_count after compacted is not billed. The compaction fixture repeats total_tokens 11081643 with a different last_token_usage (total_tokens 17983). That 17983 is no longer added. Advancing last-usage totals in that fixture sum to 390224.
• token_usage_record is a usage source, including compaction. The line is classified before the headless-usage fallback, including when "usage" sits past the 256-byte header. Before a file’s first record, usage still comes from token_count. From the first record on, only payload.usage on records is counted; later token_count lines supply quota and context only. Compaction inference that never appears on a token_count is therefore included. turn_token_usage and thread_token_usage are not summed.
• cache_write_input_tokens / cacheWriteInputTokens are an inclusive subset of input, clamped the same way as cached input.
• Record ids are kept on the parsed record only: turn_id, root_turn_id, and response_id. For a sub-agent file, session_id (the parent) and root_turn_id are available so a later change can roll cost up to the parent turn. No store columns were added.

Historical data

Stored history is corrected only when those rollout files are rescanned. CODEX_SCAN_CACHE_PARSER_REVISION is unchanged, so this does not force a full-history resync.

Tests

Synthetic fixtures under tests/fixtures/codex/usage-record/ cover a mixed-era file (legacy token_count, then records, including a compaction record) and a sub-agent file. cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, and cargo test --workspace passed.