test_cue fails on macOS (11 tests) while CI passes #138

What: tests/test_cue.py passes on macOS the same way it passes in CI.
Why: On cavallo (macOS) 11 tests in tests/test_cue.py fail on clean main eda2c8d (v1.4.0), while CI proof on Linux is green, so local proof on Macs is unreliable.
So what: Small test or platform fix for whoever picks it up.

Evidence

• Muse Spark 1.3 round-3 review of PR fix: credit interventions to their actual actor #135, run on cavallo 2026-10-19: python3 -m unittest discover -s tests at PR head 49c3766 had 12 failures, 11 in tests/test_cue.py and 1 in tests/test_health.py (Flaky test: test_hung_scan_tools_are_unavailable fails under load #136). The 11 cue failures were byte-identical on clean base eda2c8d; the PR touches neither agent_observer/cue.py nor tests/test_cue.py.
• Not reproduced by me on a Mac; reported by the reviewer with a clean-base comparison.

Elon record

• Requirements and who asked: reliable proof on every machine that runs Agent Observer (Objective: rocky, cavallo and bestia each run their own Observer); found during the 2026-10-19 review follow-up Luke requested.
• Deleted: changing Cue import behavior unless the failure shows a real macOS defect.
• Bottleneck: unknown until the failure output is read on a Mac.
• Checked myself: the reviewer's report and that the PR does not touch the cue files.

Acceptance criteria

• The cause of the 11 macOS failures is identified (test assumption versus product defect).
• tests/test_cue.py passes on cavallo and in CI.

Non-goals

• The flaky health test (Flaky test: test_hung_scan_tools_are_unavailable fails under load #136).

Blockers

None.

Required proof

• Full suite run on a Mac and in CI; independent review.