feat(rl): shared choice-aware parser/decode layer for evaluation and reward (phase 7)

[xianyu9n]

Overview

Phase 7 closes the task-level choice-to-reward contract: model outputs that speak choice ids (from PR #8) now flow through one shared, never-raising choice-aware parser that decodes them to canonical category_id before the existing evaluation / reward logic.

raw model output
  → check_stage1_choices / check_stage2_choices   (shared, never raises)
  → ChoiceParseResult.decoded  = canonical category_ids
  → existing evaluate_* / reward tables            (unchanged)

Changes

Shared choice-aware layer — src/agent/task/parser.py

• check_stage1_choices: JSON/schema → exactly 5 → unique → choice_id ∈ prompt catalog → decode to canonical category_id tuple
• check_stage2_choices: JSON/schema → answer ∈ "1".."5" (strict, no numeric coercion) → positional decode against the canonical candidates
• ChoiceParseResult: format_valid / constraint_valid / decoded / model-level output / errors, plus canonical_view() → the PR6 ParseResult contract so reward keeps a single table implementation
• Same contract style as PR6 (check_stage1/2_output): never raises on model output, programming errors still raise

Evaluation — src/agent/evaluation/classification.py

• evaluate_stage1_choices / evaluate_stage2_choices now consume the shared layer directly (removed the inline parse/decode + re-serialize round-trip); prediction is the decoded canonical result

RL reward — src/agent/training/rl/reward.py

• New reward_stage1_choices / reward_stage2_choices / reward_for_choice_result: decode via the shared layer, then the unchanged reward table — correct → 1.0, valid wrong → stage1_valid_miss 0.3 / stage2_partial 0.5, invalid / malformed → 0.0, never raises
• RewardResult gains constraint_valid (additive) so "valid but wrong" is distinguishable from "invalid"; RewardConfig and all reward values unchanged

Tests (+33)

• tests/task/test_parser_choices.py (17): valid decode, unknown/duplicate/wrong count, malformed JSON, answer 0/6/unknown, never-raise battery, canonical_view, programming errors
• tests/rl/test_rl_reward_choices.py (16): correct/wrong/invalid/malformed reward, config unchanged, and evaluation ↔ reward validity agreement on identical outputs (parametrized, both stages)

Constraints respected

• No prompt / candidate-policy / canonical ground-truth / registry / candidate-metadata changes
• No parquet re-export, no token-stats recompute, no reward-value changes
• No training run

Verification

• Full suite: 233 passed, 2 skipped (skips = verl-compat, run in CI)