flaky: bridge.test.ts hung-tree cancellation leaves the sentinel behind (~50% locally) #580

test/bridge.test.ts > BridgeClient > cancels a hung bridge tree and fails mutation closed fails roughly half of local runs on Windows:

AssertionError: promise resolved "undefined" instead of rejecting
 ❯ test/bridge.test.ts:551:37
    549|       expect(response.message).toContain("timed out");
    550|       await new Promise((done) => setTimeout(done, 1_200));
    551|       await expect(access(sentinel)).rejects.toThrow();

The test asserts the cancellation removed a sentinel file, but waits a fixed
1,200 ms for the process-tree teardown to get there. On a loaded Windows box
that wall clock is not always enough, so access(sentinel) still resolves and
the assertion reads the delay as a teardown failure.

Measured 2026-08-01 on sprint/portable-release-and-protected-state, which
touches nothing under plugins/ca-pi/: 6 consecutive direct runs of the three
PI-SEC-ADVERSARIAL suites gave 4 pass / 2 fail; a further 4 gave 1 pass /
3 fail. It surfaces through test_pi_security.py as a red PI-SEC-ADVERSARIAL
row, which is how it was found.

CI has not gone red on it so far (run 30691314813 reported PI-SEC-ADVERSARIAL
green), so this is filed rather than fixed inline.

Fix direction: replace the fixed sleep with a poll-until-gone loop bounded
by a generous timeout, so the test measures whether teardown happens rather
than whether it happens inside an arbitrary 1.2 s. Same shape as the other two
open flakes, #504 and #535.