fix(moench): Mosaic3 all-on DMD (scalar SLMImage no-op) + premature stim-mask build + napari live-restart during MDA

[hinderling]

Three Moench (Andor Mosaic3 DMD) acquisition-reliability fixes, all confirmed on the scope. The first two also exist in main, so feeding them back; the third (§3) is new.

1. fix(moench): expand scalar SLMImage to an array so the Mosaic3 actually updates

On the Mosaic3, core.setSLMPixelsTo(dev, N) is a silent no-op — no error, no effect; the previously-latched pattern stays displayed. The base MDAEngine renders a scalar-bool SLMImage (data=True/False) via setSLMPixelsTo, so every scalar SLM command was doing nothing on this hardware. Two symptoms, one root cause:

• First stim frame fired full-field (all-on). When a stim mask isn't ready, _build_stim_slm sends SLMImage(data=False); the intended all-off never reached the DMD, so the latched all-on fired the UV full-field instead of the targeted pattern.
• Occasional fully-blank imaging frame. MoenchMDAEngine._maybe_inject_dmd_wake_slm injects a scalar SLMImage(data=True) to hold the DMD all-on for each imaging capture — also a no-op, so imaging relied on the latched all-on plus the OverlapMode re-pulse and the 200 s hold. Without an active re-display each frame, the mirror occasionally parked mid-capture and the frame came back blank.

Fix: in MoenchMDAEngine._set_event_slm_image, expand a scalar-bool SLMImage to a full uint8 array before delegating, so it routes through setSLMImage (arrays are the only reliable Mosaic3 update path). This actively re-displays the intended pattern every frame, pinning the mirror against the blank-frame parking. Contained entirely in the microscope-specific engine as a hardware workaround.

Not addressed by this fix: a separate every-other-frame imaging dimming — a ~0.74x signal-only even/odd (fluorescence scales, background stays flat, in focus) that recurs intermittently on some runs — is still open and under investigation. It is distinct from the all-on / blank-frame symptoms above; do not read §1 as closing it.

Diagnostic tell for old data: a stim readout channel that is full-field (≈100% of pixels lit) is a pre-fix run; targeted thin lines mean the fix was active.

2. fix(controller): gate stim SLM build on predecessor frame acquisition

The feed loop (_run_mda_with_events) runs several events ahead of the camera (qsize<3 backpressure). In previous stim mode it built a stim event's SLM — blocking get_stim_mask(t-1) — long before frame t-1 was acquired. At minute-scale intervals with slow segmentation, the mask wait (80 s) expired before the predecessor frame even existed, so _build_stim_slm returned SLMImage(data=False) (which, combined with bug 1, fired the DMD's stale pattern).

Fix: track the (t, p) of each acquired imaging frame and, before building a stim's SLM, wait for its predecessor to be acquired — (t-1, p) in previous mode, (t, p) in current mode. get_stim_mask then only waits out segmentation of an already-acquired frame, and the 80 s timeout becomes a genuine pipeline-stuck detector again. No deadlock: the predecessor is always queued in an earlier feed-loop iteration; the wait is cancel/fatal-aware, mirroring the existing WaitEvent acquisition gate. This is pure feed-loop scheduling — the MDA engine stays microscope-agnostic.

3. fix: stop live acquisition before every MDA; drive KeepDMDAlive from live events

A third Moench acquisition-reliability fix in the same DMD / live-view area.

napari-micromanager restarts live mid-MDA → dropped/corrupted frames

With napari-micromanager attached to the same core (the usual Moench workflow), its CoreViewerLink connects configSet/exposureChanged → _restart_live, which restarts continuous acquisition whenever its live timer is set. The MDA engine sets the channel config and exposure on every frame, so a stale live timer (_live_timer_id set while the stream is already stopped — e.g. after live-picking FOVs) makes each frame restart a live stream that fights the engine's snapImage ("... sequence acquisition is running" → dropped frames). The controller already tried to stop live at MDA start but gated it on isSequenceRunning(), which misses the stale-timer case.

Fix: stop continuous acquisition unconditionally at MDA start — in both Controller._run_mda_with_events and MoenchMDAEngine.setup_sequence (covers experiments and bare mmc.mda.run() / calibration). stopSequenceAcquisition() emits sequenceAcquisitionStopped, which clears napari's stale timer, so it must run even when isSequenceRunning() is False. It runs before acquisition starts, so it only ever stops a live preview, never a real hardware sequence.

The root cause is arguably a napari-micromanager bug: _restart_live lacks the self._mmc.mda.is_running() guard that its sibling _image_snapped already has. This is the microscope-side defense; the one-line upstream fix is being reported separately.

KeepDMDAlive tied to live acquisition (positive lifecycle)

KeepDMDAlive re-displays the DMD pattern so the Mosaic3 mirrors don't park after the 200 s hold. It previously ran from boot and was stopped/restarted around every MDA — a brittle "negative" lifecycle that could race on a cancelled run (a late teardown_sequence restarting the thread after the next run's setup_sequence stopped it). It's now positively tied to live acquisition: run() on continuousSequenceAcquisitionStarted, stop() on sequenceAcquisitionStopped, with run() idempotent (repeated live-start events — e.g. napari re-arming live on config changes — can't spawn duplicate threads) and stop() a no-op when idle. During an MDA the engine drives the DMD every frame (the 200 s hold spans the inter-frame gaps), so no keep-alive is needed and there is no stop/restart bracketing to race.

Testing

• Hardware (§1/§2): confirmed on the Moench — targeted stim on the first stim frame, and stable imaging with no all-on/blank frames.
• Hardware (§3): confirmed that a stale napari live timer no longer breaks snaps (unconditional stop clears it and frames come back clean), and the KeepDMDAlive lifecycle — off at init, on during live, idempotent, off during an MDA, resumes when live restarts.
• Unit: adds tests/test_stim_gate.py, a self-contained regression test that drives the real feed loop through the fake microscope with a deliberately slow camera and asserts every stim SLM is built only after its predecessor frame is acquired (no motile/stardist dependency).