forgetest: the emission witness samples a rapid after its M5; m5-rapid-dark leaves

The rule that a rapid after an M5 ships no FIRE is the host's (laser
stream rule 16). What only the machine shows is LASER_ON staying off
with FIRE clear while the chain can fire; kernel.fire-line's dark
sample runs with the chain unarmed. laser.m5-rapid-dark held that
sample in a test of its own, with its own arm and press. The emission
witness, which every campaign runs, now holds it.

- The square ends at M5 with no program end, so the window stays open.
- The dark point: the LASER_ON count reads 0 after the M5, the window
  still armed, and the SoC lock line released (laser_locked false: the
  engine's armed flag can read true after the latch has locked, and a
  dark reading with the latch locked is the unarmed case). All three
  are checked before anything else is sent.
- Only then is the rapid sent: G0 X20, a 2.5 s dwell, G0 X-20, a 0.5 s
  dwell, the program end. Sending it after the dark point keeps the
  mid-square dwell's own zero from being taken for the dark point.
- The test samples until the window closes (a dwell reports Idle, so
  Idle cannot end it), with the controller's machine X beside each
  sample.
- Every sample from the dark point until the window closes reads no
  emission, and the HV current stays at its idle reading (HV_DARK_MAX);
  the reports show the head out at +20 mm and back at the start inside
  that stretch, and the latch reads released on the sample at +20 mm.
- The square still has to show emission and an HV rise, judged on the
  samples before the M5; the beam detector is judged there too.
- The job-based disarm is timed from the head being back: under 10 s,
  where the idle grace is 60 s.
- The rapid's list and its sampler live inside the test, and the
  deletion takes only m5-rapid-dark's own lines (decorator through
  body), so laser.py's shared text is byte-for-byte unchanged.
- live_fire_drills.py's m5dark note and the bench README say the
  emission witness now carries the drill's check.

Proof: the fingerprint parts of laser.py before and after: the shared
text unchanged, every other test's own text unchanged. In the
controller, glowforge_laser.c keeps the latch unlocked from the job's
first laser-on through M5/M3 toggles and relocks it at program end or
after the laser_disarm_s spindle-off grace (60 s by default), so the
dark point and the rapid fall inside the window. The coverage lint
reports 0 uncovered paths over 110 tests (its one failure is
setup.check-lens's src/lensfind.* entry, which the forgectrl pin bump
clears); pyflakes adds no finding; the forgetest host suite passes
under Linux.

Acceptance: only laser.emission-witness moves, and as an always-required
test it must pass on the bench in the next campaign before a release
can inherit its PASS.

scripts/bench/live_fire_drills.py

@@ -96,7 +96,8 @@ Drills (pass a name):
             reads dark after the M5, and laser_on_sampled never goes
             nonzero again after its first zero past the line. Prints the
             9 s after the line at 40 ms steps. The catalog's
-            laser.m5-rapid-dark is its port.
+            laser.emission-witness carries its check: a rapid after
+            the M5, sampled dark.
   m4corner  M4 velocity-scaled power into corners and short segments:
             a corner-heavy vector pattern (a long line, 1.1 mm zigzag
             teeth, a 2 mm square, a 180 degree reversal, 0.5 mm teeth)