Adds bead.config.compose, FORCED_CHOICE, gallery v0.4.0 wiring (0.5.0)

[aaronstevenwhite]

Description

Three coupled changes plus a wave of correctness fixes:

1. bead.config.compose — a didactic-grounded config composer
   replaces the hand-rolled loader. Generic over any dx.Model
   schema; supports the full OmegaConf interpolation grammar
   (${section.field}, ${.x} / ${..x} relative,
   ${a.b[0]} / ${a.b.0} list indexing, ${a.${b}} nested,
   \${literal} escape, cycle detection) and the standard
   resolver suite (oc.env, oc.select, oc.decode,
   oc.deprecated, oc.create, oc.dict.keys, oc.dict.values).
   defaults: [...] composition, TOML support, schema-driven
   strict-merge, and repeatable --set KEY=VALUE CLI overrides
   on every bead subcommand. Bead-specific resolvers
   (${bead.path:rel}, ${bead.anchor:name[,attr]}) live in
   bead.config.resolvers. Structured for future extraction as a
   standalone didacticonf distribution: only merge.py imports
   didactic.

2. ScaleType.FORCED_CHOICE — first-class N-AFC support
   throughout family_to_item_template, the active-learning
   model registry, and AnchorSpec.scale_type (response space =
   positional labels like ("first", "second"), per-item
   alternatives on each Item).

3. Gallery v0.4.0 wiring — gallery/eng/argument_structure/
   gets a new protocol.py module that materializes the live
   AnnotationProtocol from config.yaml. create_2afc_pairs.py
   threads the anchor name into every pair's item_metadata;
   generate_deployment.py and simulate_pipeline.py build
   their ItemTemplate via family_to_item_template instead of
   literal prompts; make validate-protocol gates make data.

Fixes

• Save/load bug in 8 model subclasses: the base trainer
  popped state-only fields from config.json before
  _load_model_components, but each subclass re-read the file
  and undid the pop. _load_model_components(load_path, config_dict)
  now takes the cleaned dict.
• ResourceWarning file-handle leaks in
  bead.cli.items_factories: for line in open(...) →
  read_text().splitlines() in 6 sites.
• ValidationError: PosixPath is not JSON serializable in
  the Lightning trainer: self.config.model_dump() →
  json.loads(self.config.model_dump_json()) so Paths / Enums
  flatten to JSON shape.
• UserWarning: participant_ids provided but mode='fixed' in
  ~40 mixed-effects tests: dropped redundant participant_ids
  args in fixed-mode tests; switched length-validation tests to
  random_intercepts so the validation path is actually
  exercised.
• spaCy 3.8.11 ↔ Python 3.14 import crash: upgraded to
  spaCy 3.8.13 / pydantic 2.13.4.

Motivation

Two coupled needs surfaced together:

• The gallery still used v0.3.0-era idioms (hard-coded prompts,
  ad-hoc JATOS-result loading, dict-shaped config). It should be
  the reference implementation for the protocol stack the
  framework now ships.
• The hand-rolled config loader (no interpolation, no defaults
  composition, no CLI overrides) was the actual blocker for any
  non-trivial declarative workflow. The cleanest answer was a
  custom composer grounded in didactic rather than pulling in
  OmegaConf and accepting a second data-structure paradigm.

The fixes were uncovered while running the full suite under
-W error to verify the new code didn't regress anything.

Fixes #

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update
• Refactoring (no functional changes)
• Tests (adding or updating tests)

Checklist

• I have read the CONTRIBUTING guidelines
• My code follows the project's style guidelines
• I have run uv run ruff check . and uv run ruff format .
• I have run uv run pyright with no errors (pre-existing pyright noise in untouched modules persists; new files are clean)
• I have added tests that prove my fix/feature works
• All tests pass (uv run pytest tests/) (3553 passed under -W error, excluding the slipt-dep tests/behavioral/ directory; the long-running gallery/eng/argument_structure/tests/test_simulation.py::TestRunSimulation BERT-training tests hang in this local environment as they did pre-branch)
• I have updated documentation as needed

Testing

• tests/config/compose/ — 43 new tests cover the parser/evaluator,
  built-in resolvers, custom-resolver registration, and the
  end-to-end compose pipeline against a FakeSchema.
• tests/protocol/test_encoding.py, tests/protocol/test_items_bridge.py
  — FORCED_CHOICE paths through encoding + items bridge.
• gallery/eng/argument_structure/tests/test_protocol.py — config
  → AnnotationProtocol round trip, scale type, prompt agreement,
  model registry routing.
• gallery/eng/argument_structure/tests/test_simulation.py —
  updated to the protocol-driven template and Item.model_validate_json.
• Full suite (minus tests/behavioral/ which requires the optional
  slopit dep): 3553 passed under -W error in ~32 minutes.

Screenshots (if applicable)

N/A