feat(usage): Capture orchestrator usage through a source ledger

[4ndreello]

codedeck open sessions recorded no usage, and they are almost all of the real spend. Worker usage also lost cost across process restarts: harnesses report a cumulative value per native session or process, and the store replaced the row's usage with the latest one.

Usage now goes through a ledger:

• usage_sources keeps a global high-water mark per source key (claude:<nativeId>#<process>, codex:<thread>, session:<id>, claude-open:<nativeId>). A new observation adds only its increase over the mark, so re-reading the same cumulative value adds nothing.
• usage_attributions records what each row got from each source and is materialized into sessions.usage_*, so existing readers keep working.
• session_native_links keeps every native id an open row saw. The session-id hook now appends ids to the sidecar, and open runs a link watcher that sends session.linkNative and flushes before finishOpenSession and session.release.
• On release, at daemon start for stale rows, and when a second id links, the daemon reads the Claude transcript (last cost-state, else deduped assistant tokens priced per model) and observes it. Release re-reads every link of the row, which covers resuming an id on a revived row and /clear then /resume inside one open.
• The statusline sends its live cost with codedeck usage <run> --json --observe <session_id>=<cost>. run $ is workers plus the orchestrator sources other than the current session plus the live payload cost.
• RunUsageSummary gains orchestrator and total; the top level stays workers only. codedeck usage gains --by origin and --backfill, which imports past orchestrator transcripts as usage_legacy entries with a mark, so a later resume of the same id counts only the increase.
• Pricing: cached tokens are subtracted from input for codex only (PRICE-01); cached above input has no known cost.
• Rows that already had usage before the upgrade are seeded into the incoming source when it has no mark, or into seed:<sessionId> otherwise (and when a Claude row's previous usage came from another process), so the upgrade neither drops nor double counts them.

Spec, design and tasks: .specs/features/orchestrator-usage/ (ORCH, RUN, SRC, PRICE, BF requirements).

The daemon must be restarted after deploy to create the new tables; the migration only adds tables and indexes.

Not covered: the Claude statusline cost after --resume and --fork-session was not checked against the real binary, Antigravity and OMP counters across processes are assumed cumulative per session, and the TUI dashboard with --by origin was not exercised. Backfilled legacy entries have no row attribution by design.

Tests

Scoped batches, one at a time, on the final HEAD:

• npx vitest run tests/orchestrator-usage-daemon.test.ts tests/usage-daemon.test.ts tests/release-interrupted.test.ts tests/session-adopt.test.ts: 56 passed
• npx vitest run tests/usage-ledger.test.ts tests/native-links.test.ts tests/usage-query.test.ts tests/usage-backfill.test.ts tests/usage-cli.test.ts tests/session-store.test.ts: 39 passed
• npx vitest run tests/pricing.test.ts tests/usage-source.test.ts tests/claude-transcript.test.ts tests/usage-schema.test.ts tests/power-database.test.ts tests/usage.test.ts: 45 passed
• npx vitest run tests/statusline.test.ts tests/link-watcher.test.ts tests/session-id-hook.test.ts tests/session-runtime.test.ts tests/run-env.test.ts tests/usage-statusline-contract.test.ts: 36 passed
• npx vitest run tests/open-action.test.ts tests/open-contract.test.ts: 54 passed
• npx vitest run tests/open-pty.test.ts tests/open-args.test.ts: 130 passed
• npx vitest run tests/power-recover.test.ts tests/run-linkage.test.ts tests/opencode-parser.test.ts tests/usage-charts.test.ts: 24 passed
• npx tsc --noEmit: clean; npm run build: ok
• codedeck usage <run> --json --observe <id>=<cost>: about 72 ms per call (62 ms before)
• Mutation probes, 13 faults: the ledger delta, the negative-delta guard, the seed rule, the PRICE-01 subtraction, codex-only cached input, first vs last cost-state, ORCH-16 previous ids, the RUN-06 exclusion, sidecar append, last sidecar id, the close-path flush and the worker-only top level were all caught. sequence <= ? vs < in the process ordinal is equivalent (event sequences are unique).

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud