Add difficulty-proportional type ratio override

[Oaklight]

Summary

• Add difficulty_weights parameter to SamplerConfig as a semantic override for type_ratios
• Allows weighting type sampling inversely proportional to accuracy (harder types get more data)
• Example: {noul: 0.5, choice: 1.0, score: 2.0} — score (~50% acc) gets 4x the data of noul (~80% acc)
• Different from per-type LR (Multi-task optimization: per-type LR, gradient balancing, adaptive loss weighting #95, didn't help) — this controls data quantity, not learning rate
• Wired through YAML config and CLI (--difficulty-weights '{"noul": 0.5, "choice": 1.0, "score": 2.0}')
• 3 new tests covering override behavior and edge cases

Closes part of #97 (difficulty-proportional sampling approach)

Test plan

• python -m pytest tests/test_sampler.py -v — 12 tests pass (9 existing + 3 new)

[clementine-oaklight]

LGTM ✅ — Clean difficulty-proportional sampling.

Component	Status
Override	✅ difficulty_weights overrides type_ratios when set
Fallback	✅ effective_ratios = difficulty_weights or type_ratios
CLI	✅ --difficulty-weights '{...}' JSON string
Tests	✅ 3 tests (override, fallback, missing type)

Addresses data quantity (vs per-type LR which only adjusted gradient scale). CI green.

[milo-oaklight]

Approved ✅

Difficulty-proportional type ratio override:

Design

• difficulty_weights overrides type_ratios for sampling
• Controls data quantity per type, not learning rate
• Inverse-accuracy weighting: harder types get more training data

Example

difficulty_weights:
  noul: 0.5     # ~80% accuracy → low weight
  choice: 1.0   # ~61% accuracy → medium weight  
  score: 2.0    # ~50% accuracy → high weight (4x noul)

Implementation

• effective_ratios = self._config.difficulty_weights or self._config.type_ratios
• Clean override pattern — no special casing scattered through code
• Used in both items_per_type allocation and window slot allocation

CLI

• --difficulty-weights '{"noul": 0.5, "choice": 1.0, "score": 2.0}'
• Parsed as JSON string

Tests (3 new)

• Override behavior (score gets ~4x noul)
• None falls back to type_ratios
• Missing type handled gracefully

CI green.

[elena-oaklight]

LGTM! Clean implementation of difficulty-proportional sampling.

Review Summary

Aspect	Status
CI	✅ All 4 checks passed
Logic	✅ Correct override mechanism
Tests	✅ 3 new tests cover key scenarios
Config	✅ YAML + CLI wiring complete

Implementation Notes

1. Override pattern is sound: effective_ratios = difficulty_weights or type_ratios — simple, explicit, no partial merge.

2. Consistent application: effective_ratios used in all 4 locations:

   • Initial type allocation (L174, L182)
   • Window slot allocation (L220, L227)

3. Missing-type handling: Test confirms graceful behavior when difficulty_weights references absent types (score in noul+choice pool).

4. CLI integration: --difficulty-weights '{"noul": 0.5, "choice": 1.0, "score": 2.0}' parses correctly via json.loads.

5. Config docs: Commented examples in all 5 YAML configs provide usage guidance without changing defaults.

Minor observation (not blocking): The docstring says "overrides type_ratios for sampling" — this is accurate for proportions but doesn't affect the window-interleaving structure itself, which is expected.