fix(acp): report cache token buckets under the anthropic_bucketed usage contract

[Yevanchen]

Fill what changed. Use N/A for irrelevant or maintainer-only items.

Summary

• Forward cachedReadTokens, cachedWriteTokens, and thoughtTokens from the ACP prompt-response usage payload in normalizePromptUsage (previously only input/output/total were kept, so cache buckets were silently dropped).
• Declare ACP usage as anthropic_bucketed instead of openai_total_with_cached_breakdown, matching OpenCode's actual token semantics (fresh input excludes cache read/write buckets).

Why

• The mosoo Cost dashboard for OpenCode (acp-fallback) agents only showed avg tokens: Cache Hit was always 0% and cache read/write spend was never billed, because the driver dropped the cache buckets OpenCode reports over ACP (opencode/src/acp/usage.ts buildUsage sends cachedReadTokens/cachedWriteTokens/thoughtTokens).
• The contract label was also wrong: openai_total_with_cached_breakdown makes the API compute billableInput = inputTokens - cacheReadTokens, but OpenCode normalizes tokens.input to exclude cache tokens ("AI SDK v6 … Always subtract cache tokens", opencode/src/session/session.ts). With cache fields forwarded under the old contract, fresh input would be under-billed; without them, cached tokens vanish from billing entirely. anthropic_bucketed is the contract that matches this payload shape (same as the Claude harness).

Verification

• Commands: bun test (1074 pass; 1 pre-existing failure in tests/acp-file-system.test.ts also fails on clean origin/main), bun run tc, bun run lint.
• Manual steps: verified OpenCode's ACP usage shape against sst/opencode source (acp/usage.ts, session/session.ts:getUsage) and the ACP schema (Usage with cachedReadTokens/cachedWriteTokens); traced the mosoo API pipeline (normalizeUsageTokens + calculateUsageCost) to confirm anthropic_bucketed prices fresh input, cache reads, and cache writes at their own rates and fixes the Cache Hit ratio (cache_read / (fresh + cache_read)).
• Not run: live OpenCode end-to-end run (test:live:opencode).

Impact

• User/API/contract changes: new usage_event/session_model_call rows for ACP runs now carry cache token buckets and the anthropic_bucketed contract; Cost dashboards start showing non-zero Cache Hit and cache-adjusted spend for OpenCode agents. Historical rows are unaffected (they recorded zero cache tokens under the old label).
• Generated files / GraphQL / DB / lockfile: N/A
• Env or config changes: N/A
• Risk and rollback: low; ACP-only payload change, API already accepts both contract values. Revert the commit to roll back.

Review

• Closest review areas: src/runtimes/acp/acp-lifecycle-events.ts (normalizePromptUsage), usage-contract handling in mosoo apps/api/src/modules/cost/domain/usage-contract.ts.
• Known trade-offs: the contract is chosen for OpenCode's semantics; if another ACP agent that reports cache-inclusive input tokens is onboarded later, contract selection may need to become per-agent. The runtime-reported cumulative session cost from usage_update is still superseded by the API's own token-based pricing at turn end (tracked separately).

[Yevanchen]

Clean-context review (delegated): APPROVE-WITH-NOTES, merge recommended.

• Bucket semantics verified end-to-end against the pinned opencode 1.18.4 source (session.ts adjusts input to EXCLUDE cache buckets) — consistent with the claude runtime's use of the same anthropic_bucketed contract; field names match exactly.
• No double counting — traced the superproject pipeline (normalizeUsageTokens → calculateUsageCost, plus the reconciliation path storing raw values with single normalization): the OLD openai_total_with_cached_breakdown label was the actual under/over-billing risk; this PR fixes it.
• Superproject consumability: anthropic_bucketed already in the UsageContract union and all emitted fields declared — no coordinated API change needed; a routine driver pin bump after merge makes it live (current pin d179ed2 predates this).
• CI green; tests cover both event paths + replay fixture.

Non-blocking follow-ups (pre-existing, this PR strictly improves data): (1) ACP thoughtTokens are recorded but never priced (claude includes thinking in output; opencode splits it out) — track reasoning-token pricing as its own item; (2) totalTokens display semantics diverge between the two runtimes under the same contract (display-only).