feat: add task-equivalence evaluation system for adapters

[Sudhendra]

Summary

Adds a comprehensive task-equivalence evaluation system that measures whether LLMs produce equivalent answers when given verbose vs compressed context. This is the correct metric for semantic compression (not perplexity/loss).

Changes

New Evaluation System (src/evaluation/)

• Multi-model validation (Claude + GPT by default)
• Task-equivalence scoring based on semantic similarity
• Artifact stripping (<think>, </tool_call>, backtick wrappers)
• Proper Qwen3 chat template handling (was broken before)
• Configurable equivalence threshold (default 0.80)

V1 Adapter Results

Evaluated models/runs/mlx/2026-01-30_17-14-36/adapter (iter-500):

Metric	Value
Pass Rate	89% (at 0.80 threshold)
Avg Compression	62% (38% token savings)
Avg Equivalence	0.859
Failures	11/100

Key Findings

• Gemini removed: Produces inconsistent scores vs Claude/GPT
• 0.80 threshold appropriate: 21/35 failures in previous runs were 0.75-0.80
• Aggressive compression fails: <50% ratio has only 26.7% pass rate

Bug Fixes

• Fixed chat template handling (was passing raw text instead of templated)
• Added --mask-prompt to MLX eval for accurate perplexity

Documentation

• Updated docs/TASKS.md with Phase 3-5 completion status
• Added docs/plans/2026-01-31-v2-production-training.md for scaling to 10K+ pairs

Testing

pytest tests/test_evaluate_adapter.py -v  # All 3 tests pass

Usage

python scripts/evaluate_adapter.py \
  --adapter-path models/runs/mlx/latest/adapter \
  --data data/training/test.jsonl \
  --limit 100 \
  --models claude gpt \
  --equivalence-threshold 0.80 \
  --include-system-prompt