[linkerlin]

Summary

失败 episode 和成功事件不再把 capture_diff_snapshot 的结果做 [:2000] 前缀切片。截断改在 helper 内完成：已跟踪 diff 会占满 2000 字时，先裁跟踪前缀，固定留 480 字给未跟踪新文件块，因此 notes.md 这类证据在大 diff 里仍然可见。预算仍是 2000，不抬高。

Failure episodes and success events no longer take a naive [:2000] prefix of capture_diff_snapshot. Truncation happens inside the helper: when the tracked diff would fill the 2000-character budget, the tracked prefix is clipped first and a fixed 480-character tail is kept for untracked new-file hunks, so a file such as notes.md stays visible. The budget stays 2000.

What changed

• capture_diff_snapshot reserves _UNTRACKED_TAIL_CHARS (480, and at most a quarter of max_chars) when the tracked diff would leave less room than that tail. The returned string is at most max_chars. If nothing eligible is untracked, the tail goes back to the tracked diff.
• _failure_diff and the success event both call capture_diff_snapshot(..., max_chars=_STORED_DIFF_CHARS) (2000). The second [:2000] slice is gone, so both arms share one policy.
• Noise skip list is unchanged: .venv, node_modules, __pycache__, .git, protected files, runtime dirs (memory/, .evolver/), and NUL binaries stay out of the untracked supplement.
• No solidify-gate, Parent-baseline, skip_validation, or val-seal changes. DEBUG #61 and DEBUG #68 residuals are untouched.
• DEBUG #57 and CHANGELOG record the follow-up.

How to test

1. uv run pytest tests/test_git_ops.py tests/gep/test_solidify.py tests/gep/test_solidify_acceptance_hook.py tests/gep/test_solidify_bench_pack.py tests/gep/test_solidify_diagnostic_funnel.py tests/gep/test_solidify_dogfood_fixes.py tests/gep/test_solidify_fitness.py tests/gep/test_solidify_helpers.py tests/gep/test_solidify_pending_soft.py tests/gep/test_solidify_soft_continue_bugs.py tests/gep/test_solidify_soft_default.py tests/gep/test_no_git_no_commit.py tests/test_auto_trigger_and_evolution_e2e.py::TestAutoEvolutionRealReliableBeneficialE2E::test_two_round_closed_loop_rollback_then_beneficial_repair -q --tb=line
   • Expected: 181 passed.
2. uv run ruff check src tests && uv run ruff format --check src tests && uv run mypy src
   • Expected: ruff clean, mypy Success: no issues found in 353 source files.
3. uv run pytest tests/ -q --tb=line -m "not slow"
   • This run: 4454 passed, 25 deselected, plus tests/gep/test_material.py::TestWatermarkIngest::test_identical_rewrite_is_noop (skipped vs rewrite_noop). That failure reproduces on unmodified master (233bb5b) in this environment and is not part of this diff.

Risk

Low. The stored diff text changes only when a tracked diff would otherwise crowd out untracked new files. The 2000-character episode/event budget is not raised. The default 80_000-character capture (policy secret scan) reserves the same 480-character tail only when the tracked diff itself exceeds that budget.

Harness/evaluator governance

Upstream governance surface: solidify stored diff scene (gep/git_ops.capture_diff_snapshot, failure _failure_diff, success diff_snapshot). Gates, Parent baseline, skip_validation, and val seal are unchanged.
Downstream EvoX impact: N/A -- snapshot text only, no bridge or contract change.
Rollout-local scope: episode/event diff string allocation inside the existing 2000-character budget.
Promotion boundary: this PR; no feature flag.
Evaluator mismatch sets: evidence (diff scene). Observation, action, repair, verification, and belief sets are unchanged.
Non-regression evidence: git_ops + solidify family (181 passed, including the two-round closed loop), ruff, and mypy. Full non-slow suite has one pre-existing test_material failure that also fails on master.
Fix-severity review: low
Owner approval: N/A
Security boundary: untracked new-file text was already eligible for the snapshot after #22. This change only keeps that text inside the truncated budget. Noise, protected, runtime, and binary paths stay excluded. No new network or secret-handling path.
Rollback: revert this commit.
Live promotion: no
Autonomous evaluator self-editing: no

Self-check

• Tests added or updated to cover the new behavior. Relevant pytest modules, ruff, and mypy passed locally. Full pytest -m "not slow" has one pre-existing failure on master (test_identical_rewrite_is_noop), noted above.
• No new runtime dependencies.
• New source files import from .py source in CI — N/A, no new modules.
• Pydantic schema factory — N/A.
• Mutable copy — N/A.
• New module-level environment variable — N/A.

Related

Follow-up to #22 / DEBUG #57. Does not close a GitHub issue.

  

[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.