fix: account for all observed requests in an ACP prompt

[jhgaylor]

Problem

PromptResponse.usage and _meta.quota.token_count report tokenUsage.last, the final model request of the turn. A tool-using prompt makes several requests; every request before the last one is dropped from the report.

Change

• Add PromptTokenUsage: captures a thread-total baseline when the prompt starts, then accumulates total - previous deltas from each thread/tokenUsage/updated notification for the active turn.
• Baseline source: latest thread total seen on the transport (covers load/resume/fork snapshots that arrive before the ACP handler exists), falling back to the in-memory session total. New sessions start at zero.
• PromptResponse.usage, _meta.quota.token_count, and _meta.quota.model_usage now carry the prompt total instead of the last request.
• New _meta.usageAccounting (version, source, scope, completeness) documents the contract.
• Plan-to-implementation continuation is one prompt: both turns accumulate into the same total.
• session/update.usage_update (context occupancy) is unchanged.

Edge cases

Input	Result
Duplicate / replayed snapshot	delta 0, no double count
Rate-limit-only notification repeating history	delta 0
Delayed notification from a previous turn	ignored (turn id mismatch)
Resume with no snapshot before first observation	first snapshot sets baseline, not charged; completeness: partial
Negative delta, non-integer, total != input + cached + output, reasoning > output (Codex counter replacement or synthetic context-full total)	delta dropped, known usage kept, partial
Cancelled turn or typed failure	observed usage retained, partial
No notifications during prompt	usage: null, partial
Native child threads, background title generation	out of scope, not counted

Behavior change for existing consumers

_meta.quota.token_count and model_usage[].token_count change meaning from last-request to prompt-total. Reports from earlier adapter versions cannot be reconstructed.

Validation

• npm run typecheck, npm run build
• vitest run --no-file-parallelism --retry=0: 563 passed, 26 skipped
• New prompt-usage-accounting.test.ts covers: multiple requests, duplicate snapshots, previous-turn replay, restored history baseline, missing baseline, missing usage, cancellation, typed failure, counter replacement, synthetic context-full total
• plan-review-events.test.ts covers plan + implementation accumulation
• Live probe, Codex 0.153.3, two model requests per prompt across three prompts including process restart/resume:

Prompt	Prompt total	Final request only
1	28,828	14,452
2	29,127	14,602
3	29,425	14,750

Follow-up commit

5845c95 removes child subagent thread entries from the transport-level usage map when the root session closes. The listener records every thread id, and closeSession only deleted the root id.