docs(artifacts): reorganize generated artifacts + docs (phase 10)

[xianyu9n]

Overview

Phase 10 tidies up artifacts/ and docs/ after the Phase 8/9 data + stats work. No training logic, prompts, parsers, rewards, parquets or algorithms changed.

• docs/ = 项目文档（设计说明、分析结论、阶段报告）
• artifacts/generated/ = 可复现生成产物（统计、对齐报告）

Changes

Reorganized: artifacts/ → artifacts/generated/

artifacts/generated/
├── alignment/                              # data_alignment_report.{json,md} (moved)
└── prompt_stats/
    ├── canonical_id_baseline/              # 旧 canonical-id 统计（legacy baseline，原样保留）
    └── choice_id/                          # choice-protocol 新统计（去 phase9_ 前缀）

Legacy stats (old prompt_token_stats_* / prompt_stats_* / prompt_token_stats_summary.json) are kept byte-identical in canonical_id_baseline/ for regressions/comparison.

New long-term docs (docs/)

• docs/design/data_contract.md — canonical category_id 数据契约（registry/corpus/target、身份策略、训练入口规则）
• docs/design/prompt_interface.md — prompt-facing choice protocol 与共享解码层
• docs/prompt_length_analysis.md — Phase 9 token 分析（old∽new p95/max、Stage1 主收益、finance/shougang 反转、worst-case 2689/32768、PromptChoice 的 decoupling 目的）
• docs/data_alignment.md — 对齐人工结论（人类维护，脚本报告在 generated/alignment/）

Moved (git rename, tracked history preserved)

• artifacts/data_alignment_report.{json,md} → artifacts/generated/alignment/
• artifacts/prompt_token_stats_* + prompt_stats_* → artifacts/generated/prompt_stats/canonical_id_baseline/
• artifacts/stage2_contract_design.md / stage3c_report.md / stage4a_report.md → docs/phase_reports/（内部路径引用同步更新）

References / default paths

• script/analysis/analyze_dataset_corpus_alignment.py: default output → artifacts/generated/alignment/
• src/agent/task/dataset_config.py: comment path
• README.md: 新增「文档与产物结构」说明（含 canonical_id_baseline / choice_id）

Not changed

No prompt / parser / reward / parquet / algorithm / training changes.

Verification

• Full suite: 233 passed, 2 skipped（与改动前一致）
• 迁移 JSON 全部可读；受控文件无旧路径残留