Flaky: queuedHighlight composer/queue classification is not settled after one event-loop turn #408

Observed

In dependency maintenance PR #407, CI run 39105399344, macOS test shard 2/2 (attempt 1) failed one assertion in tests/queuedHighlight.test.ts: a command the composer submitted and a queued one end with the same colors → the composer path colors the command word as known at line 52. That shard otherwise reported 1,465 passed, 25 skipped and one failed. Other macOS and Linux shards and Node compatibility checks passed on the same PR. A rerun was requested; outcome must be recorded before making a flakiness claim.

The test's settle() awaits a single setImmediate after a mocked asynchronous semantic classification. The test may assume more about asynchronous scheduling/classification than the contract guarantees; that is a hypothesis, not yet a confirmed root cause. Check for actual regression caused by string-width/typings before changing the test.

Acceptance

• Reproduce against current master and maintenance PR head under contention on macOS/Node 22+ and inspect exact command/semantic cache ownership.
• Determine whether production classification can actually miscolor commands or whether the test is too schedule-sensitive.
• Fix the real race if product behavior is wrong. Otherwise replace one-turn timing assumptions with bounded deterministic synchronization to the classification completion event; do not weaken the assertion or add arbitrary sleeps/retries as the only remedy.
• Prove queued and direct composer submissions agree on command-word and argument styling, including cancellations/stale updates.
• Keep tests focused and report repeated-run stability. Preserve raw shell/output semantics.

Related: PR #407 dependency maintenance, #403 separate Fish teardown race. No evidence yet that package updates introduced the failure.