…#9177)

* test(agent): reject when no matching trace arrives before the timeout

An `assertSomeTraces` / `expectSomeSpan` expectation whose span never
arrives left the returned promise unsettled: the rejection timeout only
fired when errors had accumulated, so with zero payloads the callback did
nothing and the spec hung until Mocha's 5s per-test timeout killed it with
a generic "Timeout of 5000ms exceeded" instead of failing at the intended
1s window with the real cause.

The rejection timeout now always settles the promise: the accumulated
first/aggregate error when a payload matched but assertions failed, and
otherwise a fresh error naming the timeout window so a missing span points
the reader at "no matching trace received within <timeoutMs>ms".

* test(agent): add assertNoTraces and give slow spans room after timeout reject

Rejecting the assertSomeTraces promise at its timeout broke tests that relied on
the previous never-settling behaviour:

1. "Should not be traced" cases registered an expectation that must never match,
   left the promise unawaited, and passed via a separate setTimeout(done). The
   reject now fires into that path, calling done() twice or surfacing a stray
   "No matching trace". A new assertNoTraces resolves when nothing matches before
   the timeout and rejects only when a forbidden trace arrives; the negative
   sites move to it.
2. Positive tests whose span legitimately lands after 1s (restify <5 uncaught
   exceptions ~5s, the openai/anthropic dead-baseURL retry backoff, kafka
   consumer delivery, aws-sdk cold connect, the http Node-20 timeout cases) now
   set an explicit timeoutMs. openai/anthropic instead drop retries with
   maxRetries: 0 since the endpoint is dead on purpose. The LLMObs harness gets a
   longer window because every call is a real VCR round-trip.
3. mongodb-core's BigInt test returns done() early on the version that throws
   synchronously; the assertion now registers only once the command starts, so
   no expectation is left to reject afterwards.

* test(agent): fail loudly on leaked trace expectations

An assertSomeTraces expectation that is armed but never consumed can keep its rejection timer running after the test ends. The timer then rejects a promise nobody awaits a whole timeout window later, surfacing as an unhandled rejection on an unrelated test.

1. Each expectation promise carries a cancel() that disarms its timer and drops the handler; the consumer that armed it cancels it in its own teardown.
2. The framework reset() disarms every still-armed expectation, then throws to fail the just-finished test by name. Bare subscribe() handlers have no timer and are never counted.
3. The invalid-type SDK test no longer races getEvents() against a timeout, which abandoned a poll loop that kept draining span requests and starving later tests.

* test(agent): drop settled handlers and fix the leaks the guard exposed

This PR's teardown leak guard counts any handler left in the set as armed, but
runCallbackAgainstTraces kept its handler on the timeout and rejectFirst settle
paths, so the guard rejected expectations it had already settled. Drop the
handler on every settle path.

With the guard trustworthy, three specs leaked real expectations:

1. aws-sdk sqs: timeoutMs sat on `.catch()` rather than assertSomeTraces, so the
   suppressed-consumer expectation ran on the 1000ms default and stayed armed past
   the 250ms done().
2. google-cloud-pubsub dsm and index: the DSM checkpoint expectations were never
   awaited. Register before publishing, then await, and read the expected pathway
   hash little-endian to match DataStreamsProcessor's readBigUInt64LE decode.

* test(aws-sdk): make the disabling tests deterministic

`should allow disabling a specific service` and its sqs consumer-span-kind sibling
paired two fire-and-forget assertSomeTraces expectations on a 100ms window with a
fixed setTimeout(done, 250) asserting a match counter. On a cold localstack the
enabled request's span lands after the 100ms window, so its expectation times out,
the counter stays 0, and the guard asserts 0 !== 1.

Assert the disabled request's absence with assertNoTraces and await the enabled
request's assertSomeTraces, which is what assertNoTraces documents in place of a
bare assertSomeTraces plus a separate setTimeout(done, …).

* test(claude-agent-sdk): give the agentic-call assertion room under the timeout

This PR makes assertSomeTraces reject at its timeout instead of waiting out the
mocha deadline. The full-agentic-call test registers the expectation, then spawns
the claude subprocess and runs several steps before the trace flushes, which takes
longer than the 1000ms default, so the expectation rejected before the trace ever
arrived. Opt into a 10s timeoutMs and raise the block's mocha timeout to 15s.

…#9177)

* test(agent): reject when no matching trace arrives before the timeout

An `assertSomeTraces` / `expectSomeSpan` expectation whose span never
arrives left the returned promise unsettled: the rejection timeout only
fired when errors had accumulated, so with zero payloads the callback did
nothing and the spec hung until Mocha's 5s per-test timeout killed it with
a generic "Timeout of 5000ms exceeded" instead of failing at the intended
1s window with the real cause.

The rejection timeout now always settles the promise: the accumulated
first/aggregate error when a payload matched but assertions failed, and
otherwise a fresh error naming the timeout window so a missing span points
the reader at "no matching trace received within <timeoutMs>ms".

* test(agent): add assertNoTraces and give slow spans room after timeout reject

Rejecting the assertSomeTraces promise at its timeout broke tests that relied on
the previous never-settling behaviour:

1. "Should not be traced" cases registered an expectation that must never match,
   left the promise unawaited, and passed via a separate setTimeout(done). The
   reject now fires into that path, calling done() twice or surfacing a stray
   "No matching trace". A new assertNoTraces resolves when nothing matches before
   the timeout and rejects only when a forbidden trace arrives; the negative
   sites move to it.
2. Positive tests whose span legitimately lands after 1s (restify <5 uncaught
   exceptions ~5s, the openai/anthropic dead-baseURL retry backoff, kafka
   consumer delivery, aws-sdk cold connect, the http Node-20 timeout cases) now
   set an explicit timeoutMs. openai/anthropic instead drop retries with
   maxRetries: 0 since the endpoint is dead on purpose. The LLMObs harness gets a
   longer window because every call is a real VCR round-trip.
3. mongodb-core's BigInt test returns done() early on the version that throws
   synchronously; the assertion now registers only once the command starts, so
   no expectation is left to reject afterwards.

* test(agent): fail loudly on leaked trace expectations

An assertSomeTraces expectation that is armed but never consumed can keep its rejection timer running after the test ends. The timer then rejects a promise nobody awaits a whole timeout window later, surfacing as an unhandled rejection on an unrelated test.

1. Each expectation promise carries a cancel() that disarms its timer and drops the handler; the consumer that armed it cancels it in its own teardown.
2. The framework reset() disarms every still-armed expectation, then throws to fail the just-finished test by name. Bare subscribe() handlers have no timer and are never counted.
3. The invalid-type SDK test no longer races getEvents() against a timeout, which abandoned a poll loop that kept draining span requests and starving later tests.

* test(agent): drop settled handlers and fix the leaks the guard exposed

This PR's teardown leak guard counts any handler left in the set as armed, but
runCallbackAgainstTraces kept its handler on the timeout and rejectFirst settle
paths, so the guard rejected expectations it had already settled. Drop the
handler on every settle path.

With the guard trustworthy, three specs leaked real expectations:

1. aws-sdk sqs: timeoutMs sat on `.catch()` rather than assertSomeTraces, so the
   suppressed-consumer expectation ran on the 1000ms default and stayed armed past
   the 250ms done().
2. google-cloud-pubsub dsm and index: the DSM checkpoint expectations were never
   awaited. Register before publishing, then await, and read the expected pathway
   hash little-endian to match DataStreamsProcessor's readBigUInt64LE decode.

* test(aws-sdk): make the disabling tests deterministic

`should allow disabling a specific service` and its sqs consumer-span-kind sibling
paired two fire-and-forget assertSomeTraces expectations on a 100ms window with a
fixed setTimeout(done, 250) asserting a match counter. On a cold localstack the
enabled request's span lands after the 100ms window, so its expectation times out,
the counter stays 0, and the guard asserts 0 !== 1.

Assert the disabled request's absence with assertNoTraces and await the enabled
request's assertSomeTraces, which is what assertNoTraces documents in place of a
bare assertSomeTraces plus a separate setTimeout(done, …).

* test(claude-agent-sdk): give the agentic-call assertion room under the timeout

This PR makes assertSomeTraces reject at its timeout instead of waiting out the
mocha deadline. The full-agentic-call test registers the expectation, then spawns
the claude subprocess and runs several steps before the trace flushes, which takes
longer than the 1000ms default, so the expectation rejected before the trace ever
arrived. Opt into a 10s timeoutMs and raise the block's mocha timeout to 15s.