feat(usage): opencode transcript source for the usage scorecard

[robertelee78]

Summary

The usage scorecard read claude/codex transcripts only — opencode sessions didn't exist anywhere in usage/cost attribution, the biggest remaining "second-class" remnant of the third host. opencode persists everything the index needs in ~/.local/share/opencode/opencode.db (SQLite): session rows with title/directory/parent/aggregate tokens+cost, message rows with role/tokens/cost/modelID/providerID, part rows with text/tool payloads.

src/lib/usage-opencode.mjs maps those into the index's per-session record shape (read-only via node:sqlite; malformed rows skipped, never fatal; an absent/corrupt db reads as no source, never a throw).

Two attribution rules, both grounded in the store itself:

• Cost is observed, never recomputed. Assistant rows carry opencode's own metered cost — used as costObserved on usage rows and preferred over the pricing table in aggregate and the transcript header. Rows where no message carried a cost stay null so the pricing table applies — never a fabricated $0 for kimi/openrouter/local rates ak must not invent.
• Provider is observed, never the host. The assistant row's providerID is the inference provider (provenance 'observed'). byProvider gains opencode / openrouter / local buckets; a bare host id is never presented as billing identity (ADR-0018's invariant, scorecard side).

Integration: per-session incremental cache keys (latest message time + count — a warm refresh re-parses only sessions that gained messages); carry-forward keyed on row existence; roots.opencode test seam with the same hermeticity rule as the codex ledger (overridden roots never read the real store); readSession returns the meta+turns payload through a factored sessionPayload shared with the JSONL parsers; SCHEMA_VERSION 7. Subagent sessions (parent_id) keep their own metered usage — opencode child sessions are not parent replays, so the codex double-count rule doesn't apply. Dashboard session rows badge opencode correctly with no client change (sx.host was already preferred; provider shows observed or unknown).

Verification

• 14 new tests: 8 module (usage-opencode.test.mjs) + 6 scan/aggregate/readSession integration (usage-index-opencode.test.mjs), all against fixture dbs built with the production engine; doc citations remapped to the moved lines.
• pnpm run check exit 0 (1061 kit + cjs suites), pnpm run test:surface exit 0 (25).
• Live against the real 37-session store: 35 opencode sessions indexed alongside 1547 claude + 745 codex; byHost.opencode = 35 sessions / observed cost; provider buckets opencode, openrouter, local all observed; per-session kimi metering from opencode's own rows; 3 uncosted sessions correctly fell back to the pricing table.

Scope notes

• Transcript view (/api/session/:id) works through the same parser (meta + turns). The Live tab's streaming adapters are untouched — that's a separate surface.
• OpenRouter as a provider (its session-visibility question, Usage scorecard: ingest OpenRouter session logs as a third provider #59's core) is untouched: this is the opencode-side ingestion Usage scorecard: ingest OpenRouter session logs as a third provider #59's observability axis plugs into; OpenRouter-served sessions now surface with provider: openrouter observed from the assistant rows, but no OpenRouter-side transcript parsing is claimed here.

Refs #59. Independently stackable with #85/#86/#87 (different files except usage-index.mjs ↔ none of them).