otel: read the engine's provenance flags instead of guessing from values

[brentrager]

Required follow-up to #474, closing a hole I introduced there. Consumes SmooAI/smooth-operator-core#177 (published as core 1.10.0, verified on crates.io — see below).

The hole in #474

#474 decided "was this measured?" by testing prompt_tokens > 0. That infers provenance from the value, and there are two fabrication sites in core that differ:

site	prompt_tokens
streaming	hardcoded 0
non-streaming	estimated from the outgoing request's JSON length → plausible non-zero

So the heuristic caught the streaming path (the common one, which is why #474 was still a real improvement) and silently waved the non-streaming path through as measured.

record_turn_usage now reads AgentEvent::Completed.usage_estimated. A flag carries the fact; a heuristic guesses it — the same reason absent beats 0.

Proof, by restoring the old gate and re-running the new test:

an ESTIMATED count must be omitted however plausible it looks — this is the case
the old value-based heuristic missed; span fields: {… "gen_ai.usage.input_tokens": "372" …}

372 is invented. The old gate published it as a measurement.

Two new attributes

• gen_ai.usage.cost_source = gateway | estimated, set alongside cost_usd from Completed.cost_estimated. Local ModelPricing returns the free tier for any model it doesn't recognise, so an estimate can be a wild under-count while looking exact. A billed surface must not render the two identically.
• gen_ai.response.id from Completed.response_id, recorded whenever present. Joins to LiteLLM_SpendLogs.request_id, whose row carries the gateway's authoritative dollars and real token counts — so it matters most exactly when the counts are missing, and it turns "measured vs estimated" from a claim into something reconcilable after the fact.

Both had to be declared tracing::field::Empty at span creation. Span::record silently no-ops on an undeclared field, so the first run recorded neither while everything still compiled and only the assertions caught it.

Wire compatibility

TurnUsage now carries the three engine flags, but they are telemetry-only: eventual_response still serializes exactly {costUsd, promptTokens, completionTokens}. Pinned by an assertion on the object's key count, with all three flags set to non-default values in the fixture so a leak fails the test.

TurnUsage loses Copy (response_id is an owned String); the one call site already passed by reference.

Verification

gate	result
cargo test -p smooai-smooth-operator-server -p smooai-smooth-operator	576 passed, 0 failed, 1 ignored, 46 suites (+1 new test)
cargo clippy --all-targets -- -D warnings	clean (exit 0)
cargo fmt --all --check	clean (exit 0)

Core publish verified directly rather than trusting a green workflow — literal 200 from static.crates.io for smooai-smooth-operator-core and -temporal at both 1.9.0 and 1.10.0, and the published 1.10.0 tarball was unpacked and confirmed to contain pub usage_estimated, pub response_id, ResponseId { and cost_estimated.

Cargo.lock also unifies socket2 0.5.10 → 0.6.4 for hyper-util/tokio/tokio-postgres. Both versions remain in the lock; incidental to the core bump, and I diffed the lock rather than assuming.

No changeset — the release for this repo is being driven separately.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LkCz96UfUxcai5RU4LwPnG

[changeset-bot]

⚠️ No Changeset found

Latest commit: d16d742

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

This PR includes no changesets

When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR