mazda: lateral toggle, independent TJA/MRCC, wheel HUD for lateral-engage

[mzdnick]

Problem

• invalidLkasSetting read the wrong bits and never triggered when Lane-keep Assist System was disabled in vehicle settings.
• The dash LKA button disengaged openpilot, canceling longitudinal if active.
• A physical TJA press arms MRCC on bus 0 as a side effect.
• TJA cars have no visual indication when they're in lateral-only mode.

Fix

• invalidLkasSetting keys on the intervention bits: either bit is a live setting, only both clear reads as off.
• On a TJA-enabled car the TJA button is the switch, otherwise the dash LKA button serves the same function; toggle MADS lateral.
• When MRCC was off before the TJA button is pressed, the controller cancels the MRCC state. MRCC remains active if it was already active.
• The HUD draws the steering wheel glyph while MADS lateral is ours, dimmed while arming and full once the EPS delivers.

Validation

• 620 mazda car tests + 296 safety tests host-side, ruff clean; car-wide interface, docs, platform-config and fw-fingerprint suites pass.
• Device validation pending

Note for future syncs

Event ordinals reserve @32/@33 for future jetlink sync.

Credit

Credit to @chinna244 for the TJA work, majority of the TJA button handling comes from their research and code.

AI Usage

Disclaimer: GLM-5.3 by Z.ai was used to help develop, debug, and document this submission. All changes were reviewed and validated by me.

[zephleggett]

Taken into danger-unstable for on-car testing: commit 1 (invalidLkasSetting from the CAM_SETTINGS intervention bits) as 854bb8b, merged as a union with a local fix so the upstream LANE_LINES term still counts and the setting reads off only when both bits are clear. The same branch removes the bus-2 camera TJA press (a02b4a5): route 00000105 showed it switching the car's own lane-keep setting off, so the press this PR keeps is the cause of the state commit 1 detects.

Not taken, for now:

• Commit 2: LKA_BUTTON defaults on for every platform, LANE_LINES 0 also happens on a camera fault (two of our segments) so it would emit a false LKAS press there, the new 0x440 rx check adds a disengage path for every Mazda, and the stock-long cancel change lands on every user. If resubmitted, one mechanism per PR, opt-in per platform, and the STEER_UNDELIVERED_MIN alert gate on its own since that one is a clear fix.
• Commits 3 and 4 reverse the 09-09 decision to leave the MRCC arm and the white wheel alone. Separate discussion.

[zephleggett]

Update: commits 3 and 4 are now on danger-unstable as well (74fdc86 MRCC undo, dbd88fa white wheel), cherry-picked on top of the camera-press removal; the TestCameraPress suite went with the press, the rest of the tests are intact (866 car + safety pass). Also taken from commit 2 as d088e1a: the non-delivery latch clears and stays off while the lane-keep setting is off, keyed on invalidLkasSetting. Not taken from commit 2: the nonzero-request gate on the alert (the latch zeroes the command, so the alert could never fire with it), the dash LKA toggle, the platform-wide LKA_BUTTON flag, the 0x440 rx check and the stock-long cancel change. HBC on #15 is now relay-only (60300cf): 1164 night segments show stock raising the beams with all six radar slots empty, so the far object was not what the lamps needed.

[zephleggett]

Closing: everything taken from this PR is on danger-unstable (see the comments above for what changed and what was left out). We will test on the car from a personal branch before it reaches develop.