feat(bigquery-analytics): hot-path same-session pause_orphan resolution

[caohy1988]

What

Resolves pause_orphan on the long-running TOOL_COMPLETED row (emitted when a paused tool resumes via a function_response in a user message) from the in-memory session history only.

_resolve_same_session_pause_orphan(invocation_context, function_call_id):

• False — the originating paused call is present in this session's in-memory history → definitely not an orphan.
• None — unknown / not yet settled: no id to pair on, no history, or the pause wasn't found in the (possibly trimmed) in-memory history.

The emit path stamps attributes.adk.pause_orphan only when the resolver returns a real bool; on None the key is omitted so consumers read it as SQL NULL ("not yet determined").

Why this shape (hot-path-safe split)

This is the resume path. The resolver is a zero-I/O in-memory scan: it adds no latency and introduces no session-service or BigQuery read into a plugin that is otherwise write-only.

The resolver never returns True. Declaring a true orphan requires a settled reconciliation — an off-hot-path session-service / BigQuery read after a settling delay — which is intentionally out of scope here. Adding a 30s settling sleep + a BQ read to the resume callback would be the wrong product tradeoff (resume latency) and a real architectural change (write-only producer → producer that also reads). A persistent None is exactly what that future fallback upgrades to True. Follow-up tracked separately (see #206 "settled pause-orphan reconciliation").

Tests

TestPauseOrphanSameSession: resolver returns False only on a found pause; None for not-found / empty / missing-id / no-session / id-without-matching-part; never True; and does no session_service read (an exploding session_service proves it). Two emit-path tests assert pause_orphan=False when the pause is in history and the key is omitted otherwise.

Note on local verification: I was unable to run the suite in my local environment — it's offline and the checked-out main requires a newer opentelemetry-semconv (GEN_AI_TOOL_DEFINITIONS) than is installed, and uv sync is blocked by an unrelated platform-incompatible dep (google-antigravity, no macOS-x86_64 wheel). Both changed files are syntax-validated; CI is the verification gate for the test run. Happy to iterate on any setup detail CI surfaces.

[caohy1988]

Findings: none.

I reviewed this against the locked #206 split decision. The PR keeps the resume path hot-path-safe: no added sleep, no get_session / session-service read, no BigQuery read/query, and _resolve_same_session_pause_orphan() never returns True. The only positive resolution is False when the same session history contains a matching TOOL_PAUSED origin (function_call.id present in event.long_running_tool_ids), and unknown cases omit attributes.adk.pause_orphan so consumers see SQL NULL. That matches the agreed first PR scope; the settled orphan reconciliation remains correctly deferred.

I also checked that the pause producer and resolver use the same pair key, the emit path preserves existing pause_kind / function_call_id, and the new tests cover found, not found, no session/history/id, id-without-matching-part, no session-service read, emit false, and emit omitted. git diff --check is clean.

One operational note: the PR is currently BEHIND origin/main. A merge-tree check against current main shows no conflict in the touched files, but it should be synced/rebased before final merge so required checks rerun on the current base. Local focused pytest is blocked on this machine by an unrelated OpenTelemetry semconv import mismatch in the adk-python environment, so I relied on the diff/contract checks plus GitHub checks for execution.