fix(tests): gate _FakeLoop fakes that carry probe monitor state

[dwu96]

Symptom (main breakage — blocks every open PR's backend shards)

The five terminal-subject tests #8201 added in test/test_autonudge_stop_auth.py fail on current main (b2320e02c) and on every branch cut from it, with monitor_update cannot apply legacy fields to a structured monitor (or 0 == 1 update-call counts). First seen as Backend Tests (Windows) (1) red on PR #8291; reproduced locally on a pristine origin/main worktree on Linux — it is platform-independent main breakage, not a shard flake.

Root cause: a merge race, not a bad fix

• feat: expose session monitors to agents #5184 (merged 2026-09-03 17:00Z) introduced is_structured_monitor_loop: a loop with monitor set and gate falsy is read as a structured controller record, and _monitor_update routes it to _structured_monitor_update, which rejects legacy fields (message, max_cycles, active).
• fix(autonudge): treat a merged subject as terminal when a paused loop's cap is raised #8201 branched before feat: expose session monitors to agents #5184 landed (git merge-base --is-ancestor confirms feat: expose session monitors to agents #5184 is not in fix(autonudge): treat a merged subject as terminal when a paused loop's cap is raised #8201's branch) and merged at 20:08Z with green-but-stale CI, so the semantic conflict never ran.
• fix(autonudge): treat a merged subject as terminal when a paused loop's cap is raised #8201's _FakeLoop fixtures carry probe monitor state but never set gate — a shape production never stores: the observation probe only runs on gated prompt loops (autonudge.py:2916), so any real loop with monitor.terminal_pending has gate=True. The un-gated fakes misroute into the structured branch and the legacy patch is refused.

Fix (test-only, one invariant)

_FakeLoop now sets gate = monitor is not None, mirroring the production invariant, with a comment explaining why. No src/ change; #8201's production behaviour is correct once the fixture has the shape real loops have.

Verification

• Red: the 5 tests fail on pristine origin/main (b2320e02c) locally, identical to CI.
• Green: with this change, test/test_autonudge_stop_auth.py passes 50/50 (was 45 passed + 5 failed).
• _FakeLoop is module-local (no cross-file imports); black (26.3.1) + flake8 clean.

Pattern harvest

Rule candidate: a test fake that carries a state field must also carry every invariant production couples to that field (here: probe monitor state exists only on gate=True loops) — otherwise a later dispatcher keyed on the coupled field silently reroutes the fake down a branch production never takes. Knowingly out-of-scope sibling: the general "PR merged with green-but-stale CI across a semantic conflict" class (merge-queue/re-run-on-merge policy) is a process question, not fixable in this test file.

[jeeshofone]

Independent verification, for merge confidence on this main-red:

• Reproduced on a pristine detached worktree at origin/main = b2320e02c (macOS host): 5 failed, 45 passed in test/test_autonudge_stop_auth.py, failure strings matching this PR's Symptom section exactly (four monitor_update cannot apply legacy fields to a structured monitor assertion misses + one 0 == 1 update-call count).
• Root cause concurs with the skew reading: feat: expose session monitors to agents #5184 (merged first) routes any loop whose monitor is set through _structured_monitor_update, which denies legacy fields (max_cycles) before the paused-loop branch runs; fix(autonudge): treat a merged subject as terminal when a paused loop's cap is raised #8201's five tests build _FakeLoops with monitor=_FakeMonitor(...) and patch max_cycles, so on the merged tree the deny answers before the terminal-subject precedence fix(autonudge): treat a merged subject as terminal when a paused loop's cap is raised #8201 added is ever consulted. fix(autonudge): treat a merged subject as terminal when a paused loop's cap is raised #8201's CI was green because its base predated feat: expose session monitors to agents #5184 — textually clean merge, semantic conflict.
• One safety note in favour of a test-side fix: on the structured path the legacy-field deny fires before any revival logic, so the wasted-revival hazard fix(autonudge): treat a merged subject as terminal when a paused loop's cap is raised #8201 guards against (re-arming a watch on a merged subject) cannot occur there — the deny refuses the cap raise outright. The precedence fix(autonudge): treat a merged subject as terminal when a paused loop's cap is raised #8201 encoded therefore only ever applies to legacy loops, which is exactly what gating the fakes expresses.

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of f4ec9d769e2ec4fc7141ac0c63e3697fa8643e61 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] f4ec9d7

False positive or not applicable? A repository writer can comment:
/ai-review override gpt f4ec9d769e2ec4fc7141ac0c63e3697fa8643e61: <one-sentence reason>

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed f4ec9d769e2ec4fc7141ac0c63e3697fa8643e61 — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] f4ec9d7

Verdict parsed from the review's SHA-scoped output markers for commit f4ec9d769e2ec4fc7141ac0c63e3697fa8643e61.

False positive or not applicable? A repository writer can comment:
/ai-review override fable f4ec9d769e2ec4fc7141ac0c63e3697fa8643e61: <one-sentence reason>

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of f4ec9d769e2ec4fc7141ac0c63e3697fa8643e61 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

All claims verified: is_structured_monitor_loop at autonudge.py:516-518 reads monitor and not gate as a structured record, the probe path guards on loop.gate, the five monitor-carrying fakes in this file were the only ones affected (other _FakeLoop definitions carry no monitor state), and the two tests that need explicit gate values already set .gate after construction, so the derived default conflicts with nothing.

Design-Verdict: PASS

Root-cause test-only fix at the right layer: the fixture now encodes the production invariant once, instead of patching five call sites.

[DESIGN-REVIEWED] f4ec9d7

[CrysisDeu]

Hi @dwu96 — thanks for picking this up! Heads-up that main moved again and this PR is now showing CONFLICTING, which also means its own CI can't dispatch.

This fix is currently the unblocker for the test_autonudge_stop_auth.py shard-1 red that several open PRs are inheriting through their merge refs — at least #7378, #7308, #8267, #8130, and #5539 are waiting on it to go green.

No rush at all, but whenever you get a chance, a rebase onto current main would unblock the queue. Happy to help verify on our side once it's re-pushed. 🙏

[dwu96]

Closing as superseded: #8317 (merged 2026-09-03T23:46Z) fixes the same main breakage with the same shape — gating the _FakeLoop double that carries probe monitor state — and landed first, which is also why this branch now shows a conflict (both PRs edit the same fixture lines). Verified on current main tip (5df44de2f): test/test_autonudge_stop_auth.py passes 50/50 locally.

The diagnosis this PR carried (the #8201/#5184 green-but-stale-CI merge race, the production invariant that probe state only exists on gated loops, and the independent macOS repro by @jeeshofone in the thread) matches what #8317 shipped, so nothing here is lost by closing. Thanks @bolichen97 for the fast fix.