iomgaa/Video-Tree-TRM5
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
283 Commits 2 Branches 0 Tags
441a0aa6c31515be6994be82f57e95a6cd702f1f
Commit Graph

153 Commits

Author SHA1 Message Date
iomgaa eecb86e27a feat(question_gen): add generate-v2 CLI subcommand and experiment script 
- Add generate-v2 subparser with --config, --store-dir, --db-path,
  --seed, and --dry-run arguments to tools/generate_questions.py
- Implement _run_generate_v2 async handler: config loading, video
  discovery, DI client construction, TreeIndex loading, pipeline
  invocation, and result persistence
- Add scripts/generate_questions_v2.sh following build_trees.sh
  conventions (source .env, conda run python path, MODE=mock support)
- Update app/question_gen/__init__.py to export full v2 public API:
  run_pipeline_v2, PipelineConfig, PipelineResult, QuestionFamilySpec,
  ALL_FAMILIES, CandidateQuestion, generate_one_v2, GateReport, run_gates
- Add QuestionGenStore.load_progress() for pipeline resumption
- Add integration tests for CLI help and dry-run behavior
- Update test_question_gen_api to match expanded __all__

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-12 00:04:42 -04:00
iomgaa 206c553143 feat(question_gen): add v2 pipeline with retry loop and heavy check 
- PipelineConfig: YAML-driven configuration with family_ratios, retry,
  concurrency, dedup threshold, and heavy sampling rate
- _assign_slots: deterministic round-robin slot assignment across videos
  with per-family weighted random selection
- _process_one_slot: full retry loop (generate → postprocess → gates →
  dedup) with reject-reason feedback to VLM on retry
- _heavy_check_one: blind LLM agent trial-answer for difficulty_steps
- run_pipeline_v2: orchestration with semaphore-bounded concurrency,
  progress/resume support, and store integration
- is_duplicate: cosine similarity dedup against embedding pool

Tests: 11 integration tests covering slot assignment, retry behavior,
max-retries exhaustion, full pipeline flow, progress resume, heavy
sampling, and store record completeness.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:55:15 -04:00
iomgaa cf51d2de9d feat(question_gen): add v2 generator with per-family prompt templates 
Implement generator_v2.py with:
- CandidateQuestion dataclass (canonical location)
- _load_prompt_template: loads per-family .md from store/prompts/
- _build_v2_prompt: constructs system+user messages with material context
- _parse_v2_response: JSON extraction, json_repair, field validation
- generate_one_v2: async VLM call orchestration with reject_reason support

Add 5 family-specific prompt templates:
- retrieval.md: factual recall from visible content
- reasoning.md: multi-hop inference across segments
- enumeration.md: counting/listing entities and actions
- visual.md: visual details requiring frame observation
- spatial.md: spatial relationships between objects/people

Tests: 11 unit tests covering prompt build, parse, and e2e generation.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:41:36 -04:00
iomgaa 9053233f99 fix(question_gen): gates reject_reason returns raw reason; short-circuit skips LLM 
- GateReport.reject_reason now returns gate.reason directly (no [name] prefix)
- verbatim short-circuit sets other 3 gates to SKIP without calling LLM
- test_high_verbatim_shortcircuits asserts zero LLM calls and SKIP verdicts

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:33:57 -04:00
iomgaa f74711cd11 feat(question_gen): add v2 material sampler with family constraints 
Implement sample_material_v2 module that samples tree nodes with
QuestionFamilySpec-aware constraint validation, providing richer
MaterialContext output (subtitles, cross-L2 context, frame paths).

Key components:
- AnchorContext/MaterialContext frozen dataclasses
- _validate_sampling_constraints: multi-level constraint checking
- _collect_subtitle_sentences: subtree subtitle extraction
- _collect_cross_l2_context: peer L2 event descriptions
- sample_material_v2: main entry with retry-on-constraint-violation

Tests: 11 unit tests covering normal sampling, used-node exclusion,
constraint violation retries, cross-L2 population, and subtitle
collection.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:30:56 -04:00
iomgaa 271d1682c9 feat(question_gen): add lightweight 4-gate quality check 
Implement 4 concurrent LLM-based quality gates for generated questions:
- key_verify: validates answer evidence in source material
- blind_answer: rejects questions answerable without video context
- multi_true: detects ambiguous multi-correct options
- leak_test: per-family shortcut detection (5 probe templates)

Includes run_gates orchestrator with verbatim_ratio short-circuit,
JSON response parsing with fallback, and 9 unit tests (all passing).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:30:13 -04:00
iomgaa 9627ac9cf9 fix(question_gen): raise ValueError on UPDATE of missing rows 
record_run_end, update_gates, and update_difficulty now check
cursor.rowcount after UPDATE+commit and raise ValueError if 0 rows
were affected. Prevents silent telemetry loss.

Adds three negative-path tests:
- test_record_run_end_missing_run_raises
- test_update_gates_missing_item_raises
- test_update_difficulty_missing_item_raises

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:23:48 -04:00
iomgaa 7abe92eb1c fix(question_gen): check_verbatim covers question_text + add missing blacklist patterns 
- check_verbatim now computes n-gram overlap for BOTH question_text and
  correct_option vs source texts, returning max(question_ratio, option_ratio).
  Extracted _ngram_overlap_ratio helper for reuse.

- Added 4 missing blacklist patterns: 'this segment', 'this frame',
  'the current frame', 'frame summary'.

- Added 5 new test cases covering the above changes.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:23:37 -04:00
iomgaa 6a5424a618 feat(question_gen): add SQLite run store for generation telemetry 
Implements QuestionGenStore with:
- Idempotent schema initialization (question_gen_runs + question_gen_items)
- Run lifecycle: record_run_start / record_run_end / get_run_stats
- Per-item recording: record_item / update_gates / update_difficulty
- GateReportLike Protocol for duck-type gate report compatibility
- WAL mode + foreign keys + check_same_thread=False

DDL aligns with research-wiki/schemas/question-gen-{runs,items}.md.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:15:53 -04:00
iomgaa c83d771923 feat(question_gen): add deterministic postprocess layer 
Add app/question_gen/postprocess.py with zero-LLM deterministic
post-processing for generated questions:

- shuffle_options: deterministic option permutation with answer remapping
- check_referent_blacklist: detect self-referential language (this clip, etc.)
- check_verbatim: word-level n-gram overlap ratio measurement
- has_time_anchor: timestamp and temporal phrase detection
- check_forbidden_material: T1/T7 source material validation
- run_postprocess: orchestration returning PostprocessResult

Tests: 31 unit tests covering all functions.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:14:43 -04:00
iomgaa 75e6d8c550 feat(question_gen): add 5 question family specs with sampling constraints 
Define QuestionFamilySpec, LeakTestProfile, SamplingConstraint dataclasses
and instantiate 5 families (RETRIEVAL/REASONING/ENUMERATION/VISUAL/SPATIAL)
targeting failure mechanisms M1-M5. Implement get_family_for_slot with
legal-type filtering + weighted random selection.

13 unit tests cover: full task-type coverage, skill_target uniqueness,
deterministic seeding, invalid input errors, and chi-square distribution.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:14:22 -04:00
iomgaa 811ffa648b feat(types): extend GeneratedQuestion with skill_target & difficulty_steps 
- Add skill_target (str | None) and difficulty_steps (int | None) fields
  to GeneratedQuestion dataclass with field(default=None)
- Update loader.py to pass new fields from JSON (backward-compatible)
- Update pools.py _q_to_dict/_dict_to_q for serialization compat
- Add question_gen_v2 config section to default.yaml
- Add comprehensive test coverage (7 tests)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-11 23:05:38 -04:00
iomgaa 56fe138a46 feat(tools): batch tree build orchestration with shared API semaphore 2026-07-11 11:56:36 -04:00
iomgaa 978ddef91b feat(tools): add build_trees skeleton with discovery helpers 2026-07-11 11:41:22 -04:00
iomgaa e9073bfdc2 feat(tree): expose build_async and accept injected API semaphore 
Core algorithms #1/#2/#3 unchanged: only entry wrapping and semaphore
source switch (injected vs self-built); build logic untouched.
- rename _build_async to public build_async (body unchanged)
- __init__ accepts keyword-only api_semaphore for cross-video sharing
- default path (no injection) is verbatim-equivalent to previous code
 2026-07-11 11:14:45 -04:00
iomgaa 1c21e215e2 test(tree): cover L1 entity fields and block ordering in view_node 2026-07-11 09:02:03 -04:00
iomgaa 4e0e05210d feat(search): append raw entity fields after view_node summary 
view_node 按题两轮摘要（summarize_node）会吞掉 entities/visible_text
字段信号，Agent 站在证据节点上仍漏读实体（benchmark 错题 M1，案例
786-2、872-3、750-1）。dispatcher 侧在摘要后确定性追加 [实体]/[画面
文字] 原文区块，LLM 无法吞掉。

- TreeEnvironment.node_entity_fields：按层级取 card 实体字段原文，
  去空白、去重、分号拼接；空字段省键；未知节点抛 KeyError
- _handle_view_node Phase 2.5：摘要后、子节点概览前追加实体区块
- 附带 ruff format 修正 test_tree_environment.py 两处既有格式

算法 #11（树环境语义搜索）数据访问层扩展，不改搜索算法本身。
 2026-07-11 08:53:56 -04:00
iomgaa 291a8108e1 test(agent): tighten type annotations in step-retry tests 2026-07-11 08:48:04 -04:00
iomgaa badfcce4cb fix(agent): validate non-empty retry delays; pin exhaustion semantics 
核心算法 #10（Agent Loop）：Codex 质量审查跟进，仅加固防御与
可观测，不改变循环语义。
- __init__ 校验 step_retry_delays 非空，空序列直接抛 ValueError
  （P5 fail-fast，避免首次可重试异常时 IndexError 掩盖原始 LLM
  异常、违背方法契约）
- 补耗尽语义窄测试（耗尽后原样抛出最后一次原始异常）与空序列
  构造校验测试
- run() 最终失败日志补异常类型名，便于回溯归因
 2026-07-11 08:44:47 -04:00
iomgaa e3184c11f9 feat(agent): step-level retry for transient LLM errors (20s/40s backoff) 
核心算法 #10（Agent Loop）：仅加固异常路径的韧性兜底，不改变
解析协议、hook 时序与步数语义。benchmark 错题 796-3 显示一次
SSL BAD_RECORD_MAC 穿透 GovernedLLMClient 重试栈后废掉 13 步
已积累上下文；本次在 run() Phase 1 增加步级重试（默认 2 次，
20s/40s 退避），可重试异常限定 (TimeoutError, OSError)，非可
重试异常照旧 fail-fast 整题终止，行为与现状一致（P5 显式异常）。
为满足 radon C 级复杂度约束，重试循环抽取为私有方法
_call_llm_with_step_retry，行为不变。
 2026-07-11 08:37:35 -04:00
iomgaa cf529f2c8f test(agent): add docstring to empty-content boundary test 2026-07-11 08:33:36 -04:00
iomgaa 439dc29b3b fix(agent): reject argless action in normalization; add boundary test 
核心算法 #10（Agent Loop）：修复 Codex 质量审查 Critical——
_normalize_action 仅在除 tool 外至少存在一个平铺参数键时才收拢，
{"tool": "x"} 无参结构不再被静默升级为空 args 合法结构，照旧
返回 None 走 retry 追问路径。补边界测试 + 测试辅助方法类型注解
与中文 docstring。
 2026-07-11 08:28:59 -04:00
iomgaa 6034d4172d test(question_gen): update __all__ assertion to include synthesizer exports 2026-07-11 08:21:06 -04:00
iomgaa a31b1fbf37 test(question_gen): drop tests for removed calibrate pairing validator 2026-07-11 08:19:45 -04:00
iomgaa 8d84d5e236 fix(agent): normalize fenced and flat-args LLM outputs in parser 
核心算法 #10（Agent Loop）：仅加固 _parse_response 解析路径——
剥除 ```json 围栏 + 收拢 action 平铺参数（deepseek 稳定输出变体，
案例 637-3/615-3 三连拒 0 步阵亡）。Thinking+JSON 协议、json_repair
兜底链、hook 时序与步数语义均未改动。
 2026-07-11 08:16:28 -04:00
iomgaa ba400417e2 config: concurrency=24, max_steps=40, breaker_threshold=48 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 12:27:05 -04:00
iomgaa a7ca6d15ed feat(harness): InferenceDepsRouter per-video 路由器 
按 video_id 懒加载 InferenceDeps 并缓存，路由 dispatch/prompt_builder:
- create_dispatch: 按 session_id 路由到对应视频的工具调度
- create_prompt_builder: 自动注册 qid→vid 映射并路由 prompt 构建
- 三元组 (video_id, skills_dir, prompts_dir) 缓存键

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 12:20:32 -04:00
iomgaa f21bf345a6 feat(runner): 注入 tool_dispatch_factory/prompt_builder_factory + fail-fast 
Runner.__init__ 新增 2 个可选参数:
- tool_dispatch_factory: 工具调度工厂
- prompt_builder_factory: prompt 构建工厂

infer/eval/train 模式缺少工厂时 fail-fast 抛 ValueError。
_make_tool_dispatch_fn/_make_prompt_builder 优先使用注入工厂。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 12:20:25 -04:00
iomgaa 924160c779 feat(ports): 新增 ToolDispatchFactory/PromptBuilderFactory Protocol 
4 个新 Protocol 类型:
- ToolDispatchFn: 工具调度函数签名
- ToolDispatchFactory: per-version 工具调度工厂
- PromptBuilderFn: Prompt 构建函数签名
- PromptBuilderFactory: per-version prompt 构建工厂

含 runtime_checkable isinstance 测试。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 12:20:16 -04:00
iomgaa d3be9b1322 refactor(tree): subtitle 迁入 L3Card/L2Card + 建树管线修正 
- L3Card/L2Card 新增 subtitle: str 字段（L1Card 不加）
- L3Node 移除 subtitle 字段（数据迁入 Card）
- assign_subtitles_voronoi 改写 Card.subtitle + L2 聚合
- _collect_card_strings 增加 skip_fields 排除 subtitle
- _node_full_text/_node_anchored_text 保持 字幕:/[sN] 语义
- get_subtitle 读 Card.subtitle（L2/L3）
- verify.py/synthesizer.py: l3.subtitle → l3.card.subtitle
- 迁移脚本 tools/migrate_subtitle_to_card.py（幂等，300 棵树已迁移）
- 9→6 个测试文件适配（3 个无需改动）

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 11:57:41 -04:00
iomgaa 45403b23b4 feat(repair): regenerator + supplement 防御修复 + 迁移脚本 
- 新增 app/tree/repair/regenerator.py（VLM 重生成 + 级联修复）
- supplement.py: deduplicate_field str() 防御 + inject_value strip
- patch.py: ruff format 格式化
- repair_trees.sh: conda source 激活修复
- 新增 migrate_from_trm4.sh 迁移工具
- enhance/__init__.py → repair/__init__.py 重命名

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 07:53:59 -04:00
iomgaa f57ee45dc0 fix: Codex 全量审查修正 
Critical:
- C1: assemble_mode 'plain' → 'ids'（合法枚举值）
- C2: question_id 加入 task_type slug 避免跨题型冲突

Important/Minor:
- generate_one 移除未用的 embed_fn/similarity_threshold 参数
- config.py 注释 11→12 同步
- 测试 question_id 断言更新

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 07:40:27 -04:00
iomgaa 93c9be8bfa feat(tools): generate_questions.py calibrate 子命令 
- Fisher exact test + effect size 组合判定（PASS/WARN/FAIL）
- 按 video_id 分组推理，避免跨视频树错用
- baseline 支持从 DB 读取或自动跑推理
- 对比表输出 + 退出码控制

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:47:52 -04:00
iomgaa 11f3c90200 feat(tools): generate_questions.py generate 子命令 
- VLM 出题 + embedding 去重 + 断点续跑 + 并发控制
- 单线程汇总点保证去重原子性
- 18 个单元测试覆盖 progress/exemplar/pool rebuild/JSON append

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:42:50 -04:00
iomgaa 6e46d184b8 feat(harness): add factory.py — InferenceDeps dataclass + build_inference_deps 
组装一次推理所需的全套依赖的工厂函数：
- TreeIndex 加载（FileNotFoundError if missing）
- TreeEnvironment 构建
- SkillRegistry 按需发现
- SearchToolDispatcher 装配
- PromptManager + prompt_builder 闭包

测试覆盖：正常路径、缺失树文件、skills 注入、frozen 不可变性。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:36:07 -04:00
iomgaa 5aa7cc48c5 feat(question_gen): is_duplicate + generate_one — 去重判定与单题生成编排 
- is_duplicate: 余弦相似度去重，空池短路
- generate_one: 异步重试循环，不含去重（由调用方汇总点原子执行）

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:32:39 -04:00
iomgaa 90f17e330e feat(question_gen): build_generation_prompt + parse_vlm_response 
- prompt 组装：system(角色+题型+约束+few-shot) + user(card+字幕+干扰项)
- VLM 响应解析：JSON 直接 + markdown code block 回退，四选一 schema 校验

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:28:25 -04:00
iomgaa 40b04f886e fix(question_gen): sample_anchor Codex 审查修正 
- C1/C2: Temporal Reasoning ≥3 L2 + Object Reasoning ≥2 L2 下限检查
- I1: L2 题型子帧不足时 ValueError
- I2: L3 过滤无 frame_path 的节点
- I3/I4: distractor_texts 扩展到整棵树范围
- I5-I7: 测试补强 Information Synopsis/Temporal Reasoning/Object Reasoning
- M1: Spatial Reasoning 测试断言 spatial_layout

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:24:30 -04:00
iomgaa a597a9f901 feat(question_gen): sample_anchor — 按题型层级采样锚节点 
含 6 种层级分支：L3 单帧、L2 多帧、Temporal Perception 特例、
L1 全量/采样 L2、L1-L2 混合。时间排序 + used_node_ids 排除。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:18:25 -04:00
iomgaa 9eb9b86954 feat(question_gen): AnchorContext + 12 题型-层级映射常量 
- AnchorContext frozen dataclass: 锚节点生成上下文（node_id, card_text, frame_paths, subtitle, distractor_texts）
- TaskTypeSpec frozen dataclass: 题型生成规格（level, needs_frames, frame_count, context_fields）
- TASK_TYPE_LEVEL_MAP: 12 种 Video-MME 题型 → 树层级 + 生成规格映射
- 11 项单元测试覆盖：映射完整性、值类型、层级合法性、frozen 不变性

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:10:58 -04:00
iomgaa eb15ab315e fix(config): _VIDEO_MME_TASK_TYPE_COUNT 11→12，Video-MME 实际有 12 种题型 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 05:08:05 -04:00
iomgaa 8c383354c5 fix(detector): 移除 ongoing_actions/visible_entities 空字段检测 + 新增 repair sh 
抽检确认这两个字段为空是合法内容状态（静物/黑帧/模糊帧），
VLM 重修也修不好，保留会导致断点续跑死循环。
L3 空字段检测缩减为 frame_summary + spatial_layout。
 2026-07-09 00:54:27 -04:00
iomgaa afe80a8b32 feat(repair): 断点续跑 progress 文件管理 
load_progress / save_progress(asyncio.Lock + os.replace 原子写入)
/ should_skip_video。支持并发安全的读改写和 --reaggregate-all 兜底。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 00:23:04 -04:00
iomgaa 8182cb86b1 feat(detector): 扩展空字段检测到 L2 event_description / L1 scene_summary 
断点续跑判据需要 L2/L1 层的 empty_field 检测。零 LLM 成本。
 2026-07-09 00:12:55 -04:00
iomgaa f733c13dd1 fix(llm): call_id 移入重试循环，每 attempt 独立 
消除重试时遥测主键冲突的根因。每次 attempt 独立记录，
parent_call_id 不受影响（循环外固定），更利于事后诊断重试轨迹。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 00:12:38 -04:00
iomgaa 5a91f392f0 fix(telemetry): INSERT OR IGNORE + WAL + try/except 三层防御加固 
根治遥测写入主键冲突(UNIQUE constraint)和并发写锁(database is locked)
导致的异常冒泡，遥测侧信道错误不再污染 LLM 重试链。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-07-09 00:11:54 -04:00
iomgaa d6e74f2734 feat(harness): runner.py — train loop orchestrator (#13 algorithm fidelity) 
Three-level nesting (epoch -> step -> per-skill), slow update 10-step
sequence, checkpoint/resume, early stop, probation accept/reject/rollback.

Key TRM4->TRM5 changes:
- sync -> async (all inference/diagnosis/evolve/validate awaited)
- LLMClient.from_env -> injected LLMProvider (DI via constructor)
- Direct DB/file access -> module functions (workspace/store/log)
- _TrainState as train() local, explicit param passing to helpers

Module-level pure functions extracted for testability:
resume_plan, _guard_infra_failures, _apply_batch_correctness,
_compute_total_steps, _should_early_stop, _format_applied_edits,
_fallback_summary, _write_skip_report, _outcome_to_quadrant_pairs,
_build_comparison_pairs, _batch_from_ids, _snapshot_current_skills.

Tests: 34 unit tests covering 13a-13e sub-tasks.
Radon: all functions Grade B or better.
 2026-07-07 13:43:20 -04:00
iomgaa 6baddcc17d feat(harness): validate.py — async 块序贯验证编排 + Probation 统一定义 2026-07-07 13:20:43 -04:00
iomgaa 8c9f7d2d37 feat(harness): checkpoint.py — TrainState 序列化 + 原子写 + 指纹校验 2026-07-07 13:13:37 -04:00
iomgaa 6b56fc8f98 feat(harness): inference.py — async run_inference + DI 2026-07-07 13:04:26 -04:00