fix(camera): don't reuse a pre-trigger frame in DefaultCamera.read_camera_frame

[Alpaca233]

Problem

Laser (reflection) autofocus is less stable on recent builds: it intermittently leaves FOVs defocused, most noticeably when a larger focus correction is needed.

Root cause

DefaultCamera.read_camera_frame has a fast path that returns the cached _current_frame when it is younger than exposure + strobe:

if self._current_frame and time.time() - self._current_frame.timestamp <= total_exposure_time_ms / 1000.0:
    return self._current_frame

In a trigger-then-read pattern (send_trigger() then read_frame()) this can return the frame from before the trigger, because the freshly triggered frame has not arrived yet.

The laser-AF loop is:

1. measure_displacement() — reads the spot, computes the correction (correct).
2. _move_z() — moves the piezo (correct, instant).
3. _verify_spot_alignment() — triggers a read and cross-correlates against the reference.

At step 3 the read returns the pre-move frame, so verification sees the spot still displaced, fails the cross-correlation, and move_to_target reverts the correct correction — leaving the FOV defocused.

The window governing the fast path grew from ~2.7 ms to ~27 ms once the focus camera began recomputing its strobe delay in software-trigger mode (_update_strobe_time uses the full 2064-row sensor, not the 256-row ROI), turning an unlikely race into the common case.

Fix

Only reuse the cached frame if it was captured at/after the most recent trigger:

if (
    self._current_frame
    and self._current_frame.timestamp >= self._last_trigger_timestamp
    and time.time() - self._current_frame.timestamp <= total_exposure_time_ms / 1000.0
):
    return self._current_frame

• Strobe untouched — illumination-sync behavior and imaging use of DefaultCamera are unaffected.
• Continuous / live mode unchanged — no per-read trigger is sent, so _last_trigger_timestamp stays old and every streamed frame passes the guard (fast path preserved).
• Future hardware-triggered focus camera works — send_trigger stamps _last_trigger_timestamp in both software and hardware paths.
• PhotometricsCamera.read_camera_frame has no reuse fast path and needs no change.

Testing

• pytest tests/squid/test_camera.py → 4 passed (test_read_frame_on_timeout already asserts each send_trigger→read_frame returns that trigger's own frame).
• python -m py_compile control/camera.py → OK.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR fixes a race in DefaultCamera.read_camera_frame where a cached frame could be returned immediately after a trigger, even if that cached frame was captured before the trigger (breaking trigger→read workflows like laser autofocus verification).

Changes:

• Add a trigger-timestamp guard to the cached-frame fast path so only frames captured at/after the most recent trigger can be reused.
• Expand inline comments to document why the guard is necessary and why continuous/streaming behavior remains unchanged.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Alpaca233]

[Claude Code] Fixed in commit 112c50b - reworded the comment to "at or after the most recent trigger" to match the >= condition. The >= is intentional: a frame stamped at the same tick as the trigger is the triggered frame (or later), so it's safe to reuse.