CC#1 flagged in PR #149: tests/kora_cli/handlers/test_email_
inbound_handler.py flakes under pytest-xdist (4-6 failures with
HANDLED_RECEIVED vs filtered_paused mismatch). Serial run: clean.

# Root cause

Two test files installed an OperationalStateHolder via direct
``h_mod._HOLDER = holder`` assignment without teardown reset:

  - tests/kora_cli/test_listeners/test_mcp_tools_stop_control.py
    (KR-MCP-STOP-CONTROL ST1, PR #142) — _holder_with_state helper
    mutates _HOLDER 13 times across the suite. No cleanup.
  - tests/kora_cli/test_listeners/test_mcp_audit_on_denial.py
    (KR-MCP-AUDIT-ON-DENIAL, PR #150) — success-path test
    installs an ACTIVE holder. No cleanup.

Why xdist surfaces it: xdist workers are separate processes, so
state doesn't cross workers. But WITHIN a single worker tests run
serially in scheduled order. When the scheduler put one of these
mutating tests on the same worker as test_email_inbound_handler
AND the mutating test ran first, the dirty PAUSED/ACTIVE holder
persisted. The email handler's state-gate (_check_state_gate in
email_inbound_handler.py:509) then saw the stale PAUSED holder
and returned HANDLED_FILTERED_PAUSED instead of HANDLED_RECEIVED.

Why serial doesn't flake: pytest's default test ordering puts
the handlers/ tree BEFORE the test_listeners/ tree alphabetically,
so the mutating tests never run before the email tests in serial
mode. xdist's worksteal scheduler doesn't preserve that ordering.

# Fix

Three autouse ``_reset_operational_state_holder`` fixtures (one
per affected file) that set ``h_mod._HOLDER = None`` at both
setup and teardown:

  - test_mcp_tools_stop_control.py — fixes the bleed source
  - test_mcp_audit_on_denial.py — fixes the second bleed source
  - test_email_inbound_handler.py — DEFENSIVE belt-and-suspenders
    so a future test author who forgets the teardown doesn't
    re-poison the email handler tests

The defensive autouse in test_email_inbound_handler.py is the
documented pattern: any test file whose tests assume a fresh
``OperationalStateHolder`` singleton should opt-in via this
fixture pattern, even if today's leaking-sources are all fixed.

Not changing production code (the singleton design itself stays
— spec §3 explicit non-scope). The fix is test-side only.

# Verification

Repro confirmed pre-fix: 3-4 out of 15 xdist runs failed.

Post-fix verification (xdist):
  - 30/30 xdist runs of alerts + test_listeners + handlers +
    audit + clients (627 tests) — all green
  - 5/5 explicit spec-mandated xdist runs — all green
  - Full repo (tests/kora_cli + tests/agent) xdist: 9171
    passed, 43 failed — failures all pre-existing in
    test_anthropic_adapter.py / test_backup.py /
    test_gateway_service.py / test_web_server*.py /
    test_model_switch_custom_providers.py (zero failures in
    test_email_inbound_handler.py)

Serial regression unchanged: 627/627 still green.

# Pattern for future test authors

When mutating a module-level singleton in tests:

  GOOD: monkeypatch.setattr(h_mod, "_HOLDER", holder) — pytest
        auto-restores at test teardown.

  ALSO GOOD: an autouse fixture that resets the singleton at
        setup AND teardown (the pattern landed here).

  BAD: ``h_mod._HOLDER = holder`` — the assignment persists past
        the test boundary and can bleed across the xdist worker
        into any other test in the same worker scheduled later.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>