* test(child_process): load the mock agent once per suite

The Integration and Bluebird suites called agent.load / agent.close in
beforeEach / afterEach, so every test tore down the mock agent and the next
brought a new one up on a fresh port. Async child processes spawned in a test
finish on their own 'close' event, which can fire after the test resolves —
during the teardown gap. The leaked command_execution span then flushed to the
previous test's now-closed port (ECONNREFUSED), and a stale assertion handler
matched the leftover manual 'parent' span instead, timing the next test out.
Under CI load the first miss cascaded across the whole suite.

Loading the agent once per describe block in before / after keeps a single
port alive for every test in the block, so a late span lands on the still-open
agent. The Bluebird global.Promise swap stays per-test in beforeEach / afterEach.

Fixes: https://github.com/DataDog/dd-trace-js/actions/runs/28281208791

* test(child_process): drain ls spans in span-maintenance tests

The shared mock agent now stays alive across every test in the Integration
suite. The "should maintain previous span" cases spawn an async ls child
process but never await its command_execution span, so a late flush can land
during a later test whose expectSomeSpan carries the same ls expectation and be
consumed there, masking a regression in that later command instead of failing.
Awaiting the span in the originating test consumes it before the next test runs.

Refs: #9113 (comment)

* test(child_process): load the mock agent once per suite

The Integration and Bluebird suites called agent.load / agent.close in
beforeEach / afterEach, so every test tore down the mock agent and the next
brought a new one up on a fresh port. Async child processes spawned in a test
finish on their own 'close' event, which can fire after the test resolves —
during the teardown gap. The leaked command_execution span then flushed to the
previous test's now-closed port (ECONNREFUSED), and a stale assertion handler
matched the leftover manual 'parent' span instead, timing the next test out.
Under CI load the first miss cascaded across the whole suite.

Loading the agent once per describe block in before / after keeps a single
port alive for every test in the block, so a late span lands on the still-open
agent. The Bluebird global.Promise swap stays per-test in beforeEach / afterEach.

Fixes: https://github.com/DataDog/dd-trace-js/actions/runs/28281208791

* test(child_process): drain ls spans in span-maintenance tests

The shared mock agent now stays alive across every test in the Integration
suite. The "should maintain previous span" cases spawn an async ls child
process but never await its command_execution span, so a late flush can land
during a later test whose expectSomeSpan carries the same ls expectation and be
consumed there, masking a regression in that later command instead of failing.
Awaiting the span in the originating test consumes it before the next test runs.

Refs: #9113 (comment)

* test(child_process): load the mock agent once per suite

The Integration and Bluebird suites called agent.load / agent.close in
beforeEach / afterEach, so every test tore down the mock agent and the next
brought a new one up on a fresh port. Async child processes spawned in a test
finish on their own 'close' event, which can fire after the test resolves —
during the teardown gap. The leaked command_execution span then flushed to the
previous test's now-closed port (ECONNREFUSED), and a stale assertion handler
matched the leftover manual 'parent' span instead, timing the next test out.
Under CI load the first miss cascaded across the whole suite.

Loading the agent once per describe block in before / after keeps a single
port alive for every test in the block, so a late span lands on the still-open
agent. The Bluebird global.Promise swap stays per-test in beforeEach / afterEach.

Fixes: https://github.com/DataDog/dd-trace-js/actions/runs/28281208791

* test(child_process): drain ls spans in span-maintenance tests

The shared mock agent now stays alive across every test in the Integration
suite. The "should maintain previous span" cases spawn an async ls child
process but never await its command_execution span, so a late flush can land
during a later test whose expectSomeSpan carries the same ls expectation and be
consumed there, masking a regression in that later command instead of failing.
Awaiting the span in the originating test consumes it before the next test runs.

Refs: #9113 (comment)