test: deflake pty_session wall-clock waits into predicate witnesses

[Kyzcreig]

Defect

tests/test_pty_session.py::test_reaper_loop_invokes_reap is a wall-clock flake: it starts the reaper at interval=0.01, sleeps 0.05, then asserts the reap counter >= 2. On a loaded self-hosted runner (ace-ai slot 8, run 35439419317, slice 12/16) it observed assert 1 >= 2 and went red on PR #730, which touches agent/credential_pool.py only. The stopwatch was measuring the machine, not the reaper.

Conversion

Tests only — hermes_cli/pty_session.py is untouched, no production seam was needed.

• test_reaper_loop_invokes_reap now awaits an asyncio.Event the fake reap_idle sets on its second call, under a 10s wait_for hang guard. The witness is the iteration count reached, so load costs time, never correctness.
• The four sibling drain tests in the same file replaced fixed asyncio.sleep(0.05) budgets with a wait_until(predicate) helper that polls the exact fact each test asserts (buffer contents / socket close code). Converted in one pass per wall-clock-flake-conversion — a single converted test in a required-check file keeps taxing every other PR.

RED proofs

Each mutation's red names the new witness, not an incidental error:

mutation	red
run_reaper while True → for _ in range(1)	AssertionError: reaper loop only invoked reap_idle 1 time(s)
_drain self.buffer.append(chunk) → pass	3 failures, all timed out after 10.0s waiting for the drain task to buffer ...
EOF branch stalled with sleep(3600)	timed out after 10.0s waiting for the drain task to observe EOF and close the socket

Production restored and re-verified green after each.

Load proof

scripts/run_tests.sh tests/test_pty_session.py → 9 passed.

20x loops with 64 CPU burners on 32 cores (load average 14 → 75):

• reaper test alone: 20/20 PASS
• whole file: 20/20 PASS (9 passed each run)

Upstream

The identical five assertions exist upstream at tests/hermes_cli/test_pty_session.py (our fork carries the file at tests/test_pty_session.py). Upstream PR opened separately and linked in a comment; not waiting on it.

[Kyzcreig]

Upstream port opened: NousResearch#116117 (same five assertions, upstream path tests/hermes_cli/test_pty_session.py; verified green 16/16 and both mutations RED-by-name on a clean origin/main worktree). Not waiting on it.

[Kyzcreig]

Slice 6/16 red is a pre-existing load flake, not this diff

tests/gateway/test_browser_control_broker_hardening.py::test_failed_deferred_cancel_flush_keeps_controller_offline → Failed: DID NOT RAISE ConnectionError (job 105909008161).

This PR's commit touches one file: tests/test_pty_session.py (git diff --stat fork/main..HEAD → 45 insertions, 10 deletions, that file only). Zero gateway code.

Control, run on a clean fork/main worktree (6aedf5c969) with no PR diff applied, Mac Studio 32 cores:

condition	result
idle, 10x	10/10 PASS
64 CPU burners (load avg ~13), 10x	5/10 PASS, 5/10 FAIL

Failures on the clean base reproduce CI exactly:

• test_failed_deferred_cancel_flush_keeps_controller_offline — DID NOT RAISE ConnectionError (3/10)
• test_timeout_while_disconnected_flushes_cancel_before_new_dispatch (2/10)

Cause is the same family this PR fixes: BrowserControlBroker(command_timeout=0.01) + a worker thread + thread.join(timeout=1.0) — a 10 ms wall-clock budget the deferred-cancel path must hit while the runner is busy. Filed as its own card for conversion; deliberately not widening this verified branch.

Re-running the slice.