🧪 mazda: relay stock LKAS and lane alerts - #2

The rebuilt CAM_LANEINFO decoded and re-encoded the camera's frame
through a DBC that describes 24 of its 64 bits, so 41 bits -- all of
byte 2 among them -- went out as zeros whenever our frame was on the
bus. The dash reads those bits, and on-device testing showed an
intermittent front-camera fault and a missing engaged departure display.
Relay the camera's exact bytes instead, on each new camera frame, with
only the three hands-warn bits masked -- and only while steering.

CAM_LKAS now carries the camera's LDW and LINE_NOT_VISIBLE bits so an
engaged departure still reaches the dash, and its counter continues the
camera's sequence at each engage edge instead of jumping by an arbitrary
phase.

…arning

Device testing of the byte-exact 0x440 relay still showed the v1
wrong-side signature engaged (right departure -> left flash, left ->
nothing), so the dash reads departure-side bits from 0x243 as well. The
engaged 0x243 is now an overlay on the camera's exact frame: openpilot
writes only the torque field, the counter and the zero-angle pattern,
adjusting the checksum by exactly the fields touched (a delta off the
camera's own checksum, so bits outside the formula's model keep the
camera's own contributions). Every other bit, defined or not, is the
camera's.

The dash hands warning is the camera's in both states now: relaying
openpilot's steerRequired events painted an orange steering wheel for
every wheel-touch and distraction alert.

The steer_required parameter on the HUD relay was dead once the camera
owned the hands warning; the masks, their tests and the stale comments
went with it. cam_laneinfo was a dead write -- the FSC settle gate reads
the parser directly -- and the overlay now derives the counter and
torque for its checksum delta from the curated bytes instead of taking
them as parameters. The FRAME_RAW container signals gain CM_ comments,
since no other DBC uses that construct.

v4 on-device: engaged lane-departure warnings are correct on both sides,
but openpilot could only steer while the camera saw lanes. The EPS gates
torque on the camera's line-visibility state in 0x243, and the overlay
relayed it -- the curated build had always forced it off ("they all work
just fine if set to zero"). The overlay now clears the bit and the
checksum delta pays for it; LDW and the undocumented side bits still
ride through, so the dash alerts stay correct. The curated fallback
returns to upstream's zeros for both bits.

v5 on-device: steering and lane-departure warnings are correct in both
states. Two gaps left. The camera's hands warning, relayed byte-exact
while engaged, tracks "LAS applying torque" rather than the driver, so
the orange wheel was on nearly whenever lane lines were drawn. And with
the steerRequired mapping deleted, openpilot alerts -- the turn-limit
warning above all -- lost their dash channel, which the stock setup had
always given them.

While openpilot steers, the three hands-warn bits now carry openpilot's
hold-the-wheel alert: cleared when quiet, set when steerRequired is up
(the pre-branch behavior). While not steering, the camera's own warning
passes through untouched.

The LAS visuals were a side effect of the passthrough: while engaged,
the camera's lane lines drew on the dash and HUD alongside openpilot's
own screen. Per the keep-the-alerts tradeoff, the relay now blanks
LANE_LINES (LKAS disabled) whenever openpilot steers without an alert
of its own, and relays the frame byte-exact while a hold-the-wheel
alert is up -- the turn-limit warning keeps the rendering path the car
already knows, lines and wheel together. Not steering, the camera's
frame passes through untouched, flags-up windows included.

The 2 Hz hold grid keyed off the controller frame counter, whose phase
has no relation to when the camera went quiet. The first hold could
land up to 0.5 s past the 1 s stale window, so the dash briefly lost
its lane display on a camera dropout. Key the grid to
laneinfo_age_frames instead: the first hold fires 1 s after the
camera's last frame, then every 0.5 s.

Comment audit: 26 lines to 9. Each invariant is now stated once, at
the place a reader needs it. The EPS-gates-torque fact appears once
(at the write masks) instead of three times; the LAS-vs-driver story
lives at the call site; fix history and caller behavior move out of
mazdacan. No code change.

The orange steering wheel is not a hands-on-wheel icon: in stock it
lights while the LAS corrects back to the lane, the EPS applying
torque. The DBC's HANDS_* signal names misled the relay's naming and
comments. The parameter becomes steer_indicator, the masks
STEER_IND_B6/B7, the comments state the observed meaning, and the DBC
signals carry CM_ notes recording it. Also picks up the stale-hold
watchdog switching to the age counter and trimmed comments.