fix: register the four firmware error codes the table was missing

[X-F-R]

Summary

Parsing the firmware headers turned up four error codes this SDK cannot name, and one
entry that describes behaviour the firmware changed on 2026-09-24. The two protocol-sync
suites read a real litearm-stm32 checkout and skip silently without one, so none of
this was visible until a checkout was placed next to the repo.

Same firmware tree as #19; this is the error-table half rather than the cartesian half.

Changes

• errors.py — register four codes the firmware emits:
  • (0x21, 0x02) speed percent out of range (usb_cmd.c:637)
  • (0x31, 0x01) preset payload too short (usb_cmd.c:908)
  • (0x39, 0x02) q contains NaN/Inf (usb_cmd.c:1034)
  • (0x02, 0x06) a move_p late terminal reply (kin_runner.c:173)
• errors.py — (0x31, 0x02) read "preset invalid or payload too short". The
  firmware split those apart on 2026-09-24, so the "or" makes a caller read a length
  error as a preset error.
• tests/test_error_codes.py — the symbolic-site manifest counted usb_cmd.c only,
  while the parser walks the whole firmware tree, so the kin_runner.c site had never
  been seen. Manifest 14 -> 15.
• tests/test_error_codes.py — an assertion claimed MOVE_P has no drop_hold gate. That
  covers the dispatch path only: the reply path (kin_runner_flush_move ->
  ctrl_accept_move_j) does have the gate and does return 0x06, so a caller could
  misread a drop-hold latch. Replaced by a positive assertion.
• tests/test_protocol_sync.py — test_bench_model_axis_matches_sdk_constant used
  #if\s+LITEARM_BENCH_1J(.*?)#else with re.S, which hits the literal #if and
  #else inside the file's own comment, so it reported "bench branch not found" for the
  wrong reason. It now matches the #define line. The firmware removed bench mode on
  2026-09-28 and left an #error guard, so the test asserts the guard is present and
  that the constant is still carried for 1J firmware.
• _protocol.py — mark BENCH_MODEL_AXIS LEGACY with the removal rationale.

Testing

$ python3 -m pytest -q
657 passed, 2 skipped                    # the 2 skips are the LITEARM_LIVE tests

$ LITEARM_LIVE=1 python3 -m pytest -q
659 passed

Before this branch, same firmware checkout, from main:

$ LITEARM_LIVE=1 python3 -m pytest -q --tb=no
3 failed, 655 passed

The three failures are exactly the ones this fixes:
test_every_firmware_literal_err_code_is_registered,
test_symbolic_site_manifest_matches_firmware,
test_bench_model_axis_matches_sdk_constant.

The 2 skips are expected offline: the protocol-sync suites call pytest.skip when no
firmware checkout is present, by design.

Issues

Refs #19

[github-actions]

🎉 This PR is included in version 1.1.1 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀