v0.1.3

Fixed

• Job error attribution race: after a GRBL error halted a job, the
  emergency cancel's safety shutdown could race into the alarm state and
  overwrite the recorded error with a secondary error:9 ("Commands
  Locked"). The first error recorded now wins, so consumers reliably see
  the root-cause failure instead of a teardown side effect.
• Secondary errors during teardown no longer break recovery: they still
  resolve pending interactive command futures and still trigger the job
  halt — only the redundant state mutation is skipped.
• test_error_halts_stream was flaky on slow runners: an emergency soft
  reset racing in-flight motion legitimately leaves the machine
  alarm-locked (real Grbl behaves the same). The test now recovers with
  $X and waits for the status poller to report Idle before starting
  the next job.

Added

• init_logging(filter, path) binding: Rust-side session diagnostics
  (stall detection, liveness, unexpected device lines) now reach stderr
  or a dedicated log file instead of being silently dropped when no
  logger is installed. The raydriver run CLI gains --log-file.
• Terminal job errors are published through the device state
  (state.error): run() resolves normally even on failure, so the
  state is the only way consumers can detect a failed job. Protocol
  errors keep their proper GRBL error codes; a stale error is cleared
  on the next job start.
• Newline-terminated, rate-limited progress output when stdout is not a
  TTY, so redirected log output never interleaves with the progress
  line.

Tooling

• Type checking is now enforced: pyright runs over python/ and
  tests/ in CI, and a .pre-commit-config.yaml (ruff + pyright) was
  added. Generated stubs were fixed along the way — parser functions
  previously landed in the root stub, where they do not exist at
  runtime, leaving raydriver.grbl.parser untyped.
• CI simplified to a plain four-platform test matrix
  (ubuntu, windows, macos-15-intel, macos-15).

Full changelog: v0.1.2...v0.1.3