fix: aggregate OpenCode usage across completed steps

[RobertTLange]

OpenCode emits per-step usage, but Headless previously retained only the last step. Two completed steps costing $1 and $2 now report $3, with all input, output, reasoning, and cache tokens included.

Completed steps are deduplicated by session/message/part identity and retained independently of the rolling text trace, so long --json --usage runs preserve earlier charges. The accumulator retains numeric usage and bounded identity data. Missing native costs fall back to estimated pricing for all recorded tokens, including separate reasoning output. The public usage schema is unchanged.

Validation:

• Regression tests reproduce last-step-only totals and eviction after more than 256 KiB of streamed output; both now pass.
• Tests cover repeated snapshots, anonymous steps, missing/invalid costs, zero usage, cache/reasoning tokens, and oversized identities.
• Focused parser/CLI tests, TypeScript build, and package check pass. Full local npm run check passes with an isolated temporary HOME and umask 0077; PR CI is checked before merge.
• Downstream agentic-scientist regression confirms the normalized total imports exactly once and repeated ingestion is idempotent.

DeepSeek tariff updates and interrupted-run recovery are separate follow-ups.

A separate test-only commit makes the existing run-wait transcript fixture timestamp deterministic. Filesystem timestamp precision could otherwise make the fixture appear older than its run and hang the suite.