[linkerlin]

Summary

失败现场和成功路径共用的 diff 快照以前只跑 git diff HEAD，宿主新建且未 git add 的文件（例如 notes.md）不会进入 episode。现在这份快照会把未跟踪新文件收成 diff --git 新文件块，两臂口径一致。

Failure and success snapshots share capture_diff_snapshot. It used to run only git diff HEAD, so an untracked new file such as notes.md was missing from the episode. The helper now appends those files as new-file hunks. Both arms stay on that one helper.

What changed

• capture_diff_snapshot 仍先取 git diff HEAD（已跟踪文件的暂存与未暂存改动，含已经 git add 的新文件），再追加从未入库的未跟踪新文件。

• 补充块跳过爆炸半径不计的路径（.venv、node_modules、__pycache__、.git、.env 等受保护文件）、运行态目录（memory/、.evolver/ 等，与 _is_runtime_state 同口径），以及文件头含 NUL 的二进制。读入量不超过快照剩余预算。

• 失败侧 _failure_diff 与成功侧事件的 diff_snapshot 都走这个 helper。未改固化门、不写 Parent baseline、不调用 skip_validation、不开 val 密封。未动 DEBUG #61，也未重开未完成 run 的 tick 宿主契约。

• DEBUG #57 标为已修。回归：未跟踪 notes.md 在 hypothesis_missing 的 episode 里可见，成功事件的 diff_snapshot 里同样可见；噪声与二进制不出现。

• capture_diff_snapshot still starts with git diff HEAD (staged and unstaged edits to tracked paths, including a new file already git added), then appends never-added files.

• The supplement skips blast-radius noise (.venv, node_modules, __pycache__, .git, protected files such as .env), runtime-state dirs (memory/, .evolver/, same heads as _is_runtime_state), and binaries whose first bytes contain NUL. Reads stay within the remaining snapshot budget.

• _failure_diff and the success-path diff_snapshot both use this helper. No solidify-gate, Parent-baseline, skip_validation, or val-seal changes. DEBUG #61 and the unfinished-run tick host contract stay deferred.

• DEBUG #57 is marked fixed. Regression: untracked notes.md is visible on hypothesis_missing and on the success event; noise and binaries are not.

How to test

1. uv run pytest tests/test_git_ops.py tests/gep/test_solidify.py tests/gep/test_solidify_soft_continue_bugs.py tests/gep/test_solidify_fitness.py tests/gep/test_solidify_acceptance_hook.py tests/gep/test_solidify_soft_default.py tests/gep/test_no_git_no_commit.py tests/gep/test_sprint23_bandit_novelty.py tests/gep/test_sprint23_coverage_gaps.py tests/gep/test_policy_check.py -q
   Expected: 166 passed.
2. uv run pytest tests/test_auto_trigger_and_evolution_e2e.py::TestAutoEvolutionRealReliableBeneficialE2E::test_two_round_closed_loop_rollback_then_beneficial_repair -q
   Expected: 1 passed.
3. uv run ruff check src/evolver/gep/git_ops.py tests/test_git_ops.py tests/gep/test_solidify.py
uv run ruff format --check src/evolver/gep/git_ops.py tests/test_git_ops.py tests/gep/test_solidify.py
uv run mypy src/evolver/gep/git_ops.py
   Expected: ruff clean, mypy Success: no issues found in 1 source file.

Risk

Low. Recorded diff text can now include untracked new-file hunks, so the existing secret scan on this snapshot sees new source files, and later novelty comparisons can see their + lines. Tracked diffs stay first and are still truncated the same way. Noise trees are not appended. Accept/reject gates are unchanged.

Harness/evaluator governance

Upstream governance surface: gep/git_ops.capture_diff_snapshot (the diff scene stored by solidify failure and success events). Gate logic, Parent baseline writers, and val seal are untouched.
Downstream EvoX impact: N/A — no bridge or contract change. Recorded diff text may contain untracked new-file hunks.
Rollout-local scope: snapshot contents only; no proposal, shadow, or cohort change.
Promotion boundary: N/A — not a gate or default-policy promotion.
Evaluator mismatch sets: evidence (the stored diff). Observation, action, repair, verification, and belief sets are unchanged.
Non-regression evidence: untracked-file unit tests on the helper, hypothesis_missing episode, and success diff_snapshot; 166 related tests plus the two-round closed-loop e2e.
Fix-severity review: low
Owner approval: N/A
Security boundary: untracked source can now appear in the diff already scanned for secrets. .env and other protected paths stay out of the untracked supplement. No new network or host tool.
Rollback: revert this commit. Snapshots return to git diff HEAD only.
Live promotion: no
Autonomous evaluator self-editing: no

Self-check

• Tests added or updated to cover the new behavior; gates pass locally: uv run pytest (relevant modules, 166 passed, plus 1 closed-loop e2e), uv run ruff check / ruff format --check on touched files, uv run mypy src/evolver/gep/git_ops.py.
• No new runtime dependencies without a clear justification in "What changed".

Related

DEBUG #57

  

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.