feat(core): execution metadata on the durable log + foldTurnRecord

[lipowen]

What

Two small, stacked additions to the agent core — the shared foundation for the observability and eval lines:

1. Execution metadata on the durable log (usage + durationMs)

• ModelDelta's terminal done gains usage?: ModelUsage — { inputTokens, outputTokens, raw? }, normalized across providers, the provider's own usage object preserved as raw (mirroring ModelFinish.raw). Any adapter can claim it; no claim = no field (spread, never assigned undefined, so the pre-existing delta shape stays byte-identical).
• The engine persists the claim on the assistant Msg at commit time, together with a new durationMs (model-call wall-clock, measured before the commit so storage time never inflates it). Every executed tool Msg records its own durationMs; the synthetic results a cancelled batch commits carry none — nothing ran. On a crash-replay the cached step skips, so the persisted numbers honestly describe the run that actually paid.
• The messages log is the one durable surface a fold can read back (step checkpoints are presence-checked, never enumerated), so this lands cost/latency accounting with no side table and no schema change (stores serialize Msg as JSON).
• The anthropic adapter previously dropped the SDK's usage on the floor; finalMessage().usage now maps through usageFromAnthropic (exported beside finishFromStopReason; partial claims are dropped whole — no half-truth for a cost report to trust). replyStream accepts an optional third usage argument for scripted models.

2. foldTurnRecord(msgs, turnId) + AgentSession.record(turnId)

The durable log already IS a complete account of a turn — this fold assembles ONE turn's interleaved rows into a self-contained record: opening (user text or attributed proactive trigger), steps (model steps with usage/duration; tool steps with the input recovered from the requesting model step, so the record stands alone), terminal status by the live engine's own condition, final text, and an aggregate usage that exists only when every model step claimed it (a cost report can't silently undercount). Anything the log cannot distinguish (failed / cancelled / suspended — live-only outcomes) folds to "incomplete", never guessed.

Why

This record is the atom the observability and eval lines share:

• a trace exporter folds on turn-terminal and ships one document (OTel GenAI-shaped, in a follow-up);
• an eval dataset curates records from production and replays them against the frozen tool results — June's durable-by-construction turns make "prod trace → eval case" structurally free, which most agent frameworks have to assemble by hand;
• feedback (a future recordFeedback(turnId, …)) anchors on the same record.

Neither change touches the server↔core contract shape (additive optional fields; the server constructs none of them), so no RUNTIME_API_VERSION bump.

Tests

• adapter: usageFromAnthropic mapping (full claim + raw passthrough; partial/absent dropped whole), done-delta carries normalized usage, byte-identical shape when the transport omits it (existing tests unchanged).
• engine: usage + durationMs land on assistant/tool Msgs; no-claim assistant rows carry no own usage property; a cancelled batch's synthetic results carry no durationMs.
• fold: completed tool-loop turn (steps order, standalone tool input recovery, aggregate usage), incomplete turn after a mid-turn crash (no text, no aggregate), unknown turnId → undefined, proactive opening attribution.

Full suite: 1181 pass / 0 fail.

https://claude.ai/code/session_01CQ4hXUQHnkRPfPDSdQZsEx

[lipowen]

Two independent review passes (correctness/durability + API design) ran against this branch; findings applied in bdac6fa:

• ModelUsage semantic contract pinned — inputTokens = TOTAL input consumed, cache reads/writes included. Anthropic's input_tokens excludes its separately-reported cache fields (unlike OpenAI/Gemini headline counts), so the adapter sums them back in and surfaces the split via normalized cachedInputTokens/cacheCreationInputTokens. Cross-provider cost sums no longer undercount cache-heavy turns.
• Sync local tools record no durationMs — no I/O means workerd's request-frozen clock would read 0; recording nothing beats recording a lie. Remote (async) tools keep their wall-clock.
• AgentSession.record() → turnRecord() — noun getter aligned with transcript()/snapshot(); record reads as a write verb and channels already use feedback.record for one.
• Aggregate-usage honesty rule now pinned by a test (one unclaiming model step drops the aggregate, per-step stays).
• Changesets bumped to minor (new public API, matching the google-drive precedent).

What the reviews confirmed correct and unchanged: exactly-once/replay determinism, byte-shape spread discipline, the completed condition's equivalence with foldEvents/result(), no RUNTIME_API_VERSION bump needed (server treats Msg as an opaque JSON blob). Full suite: 1183 pass / 0 fail.