feat(metrics): surface turn-level spend on turn_metrics and /v1/admin/metrics

[yzxcj797]

Summary

Follow-up to #586's second half: the budget now debits the LlmCallUsage the harnesses meter (#592, #638), but an operator planning which models to run qm on still has to join spend out of session_llm_requests by hand — turn_metrics carried the cache columns and no cost, no output tokens. This lands turn-level spend where the rest of the turn telemetry already lives.

Implementation

• Harness turn results gain costUsage { outputTokens, costUsd } alongside the existing cacheUsage:
  • pi sums its per-call stats (sumCostUsage next to sumCacheUsage, both return sites);
  • claude reports the SDK's running total_cost_usd delta with usageTotals.output; the finalResult fallback branch keeps output tokens but reports cost 0 — the SDK result carries no cost total, and turnsWithKnownCost distinguishes that downstream;
  • mock reports its deterministic usage so tests observe the pipeline.
• The orchestrator lands both fields on TurnMetricSample; turn_metrics grows output_tokens BIGINT and cost_usd DOUBLE PRECISION via the sink's ALTER TABLE ADD COLUMN IF NOT EXISTS migration, so existing deployments migrate in place.
• GET /v1/admin/metrics returns a spend block: samples, turnsWithKnownCost, costUsdTotal, outputTokensTotal — the scope-level number an operator plans against, next to the existing cache block.

Testing

• New test/turn-metrics-cost.test.ts drives a mock-harness turn through the full app and asserts the spend block on the metrics endpoint, plus the empty-scope shape. Differential: both tests fail with the source changes stashed (no spend block).
• Neighbors green (metrics-sink, cache-observability, budget: 18/18); tsc --noEmit clean.

Follow-up to #586

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}