- Remove AssemblerMeta.resolve_params in favor of Step.get_assembler_kwargs()
- Replace step.assembler_name/assembler_params with ASSEMBLER_NAME ClassVar
  and real typed attributes on each Step subclass
- Add MachineDefaults dataclass wired through the pipeline
- Thread Step through pipeline instead of separate assembler_name/params
- Add AssemblerRegistry with _register_laser_assemblers fallback for workers
- Make task_mgr a required parameter on AddonManager, passed at construction
  so _build_and_update_manifest always finds shared_state
- Remove set_task_manager method (redundant with constructor parameter)
- Remove _ensure_assemblers_registered conftest fixture (assembler registry
  fallback handles it)
- Step.from_dict dispatches via step_registry; workers get addon manifest
  through shared_state automatically
- Remove _STEP_TYPE_TO_ASSEMBLER hack from compute_workpiece_artifact
- Add to_dict/from_dict roundtrip and get_assembler_kwargs tests for all
  step subclasses
- Clean up pipeline test args indices (args[7] -> args[5])
- Remove addon imports from core pipeline tests

- Add orchestration ClassVars to Step base (IS_VECTOR, SPLIT_CONTOURS,
  SET_POWER, NORMALIZE_WINDINGS, ALWAYS_WRAP, SECTION_TYPE) and
  override on each Step subclass - replaces AssemblerMeta fields
- Delete AssemblerMeta class entirely; AssemblerRegistry becomes a
  plain name -> Callable map with no param_keys filtering
- Add set_step_property() helper to StepComponentSettingsWidget base
  for undoable step attribute writes via ChangePropertyCommand
- Migrate contour, frame, shrinkwrap, wavefront, and raster widgets
  to read/write step attributes directly instead of opsproducer_dict
- Add override_threshold, threshold to ContourStep and dither_algorithm
  to EngraveStep (previously only on producers)
- Fix Step.from_dict to delegate to step_class.from_dict(data) so
  subclass attributes are deserialized correctly
- Remove _register_laser_assemblers() hardcoded fallback from registry
  (consistent with ProducerRegistry/StepRegistry pattern)
- Simplify all assembler registration calls (no AssemblerMeta wrapper)
- ShrinkWrapStep imports prepare_surface directly from rayforge.image
- Fix UnboundLocalError from local logger shadowing module-level logger

- Delete all producer files: OpsProducer, ProducerRegistry, PlaceholderProducer,
  and all 6 laser producer implementations (contour, frame, shrinkwrap,
  raster, wavefront, material_test_grid)
- Remove register_producers hook spec, registration table entry, and all
  implementations (laser_worker, sketcher)
- Move CutSide/CutOrder enums to rayforge/core/cut_side.py
- Migrate material test widget + command to step attributes (no more
  opsproducer_dict or producer parameters)
- Widget frontend passes 4 args (editor, title, page, step) — no more
  producer arg
- Base class makes component parameter optional with default None
- Add missing_step_types property to Doc (replaces missing_producer_types)
- Clean up all remaining OpsProducer/opsproducer_dict/producer_registry
  references across tests, scripts, and hooks

Fixed-corner mechanical/firmware skew between left-to-right and
right-to-left raster passes misaligns alternate scan rows on some
machines, regardless of speed (a different case from the
timing-delay skew LightBurn's speed table addresses).

- Adds a bidir_x_offset_mm param to the Rasterizer producer (UI
  control + serialization), per engrave step rather than a single
  machine-wide value, since different jobs may want different
  corrections, and this may need to vary by speed in the future
- Adds BidirScanOffsetTransformer, which shifts each right-to-left
  raster pass along X by that offset during gcode generation,
  preserving per-pixel power values for variable-power (photo)
  engraving
- Tests the producer -> settings -> transformer wiring end-to-end,
  not just each piece in isolation

- Doubles the auto-calculated overscan distance for individual test
  blocks, so backlash settling happens outside the visible engrave
  area rather than muddying whatever parameter the grid is testing
- Adds Optimize to the default transformers, but disabled: its
  nearest-neighbor travel reordering has no concept of cell
  boundaries and was interleaving lines from different cells
  instead of engraving each one fully before moving to the next.
  Left toggleable rather than removed, so it's easy to compare
  with/without

Extends the material test grid so users can calibrate bidir_x_offset_mm
empirically: columns vary speed, rows vary offset. raygeo's
generate_material_test_grid already supports grid_mode="Speed vs
Offset" plus min_offset/max_offset (including scaling power with speed
so darkness stays comparable across a row) -- this just plumbs an
offset_range param through the producer, its preview, and the step
settings widget.

Add bidirectional raster scan X-offset calibration

Re-implement bidirectional raster scan X-offset calibration feature
from main in the new assembler/step architecture (producers were
deleted in this branch):

- EngraveStep: add bidir_x_offset_mm attribute with serialization
- MaterialTestStep: add offset_range attribute, pass min/max_offset to
  assembler, add SPEED_VS_OFFSET grid mode support
- material_test_helpers: add SPEED_VS_OFFSET to GridMode enum, pass
  min/max_offset to draw_preview
- material_test_cmd: add offset_range to _PARAM_KEYS for preview sync
- raster_widget: read bidir_x_offset_mm from self.step (not producer)
- material_test_grid_widget: read offset_range from self.step, fix
  offset handlers to use tuple signature of _update_range_param
- workpiece_stage: read bidir_x_offset_mm from step.to_dict() via
  setdefault instead of deleted opsproducer_dict
- Add BidirScanOffsetTransformer and tests from main (unchanged)
- Update tests for new architecture (no opsproducer_dict references)

…peline

refactor: replace OpsProducer pipeline with assembler registry