Implementation Plan: Add Deterministically Gated Overengineering Auditors to review-pr

[Trecek]

Summary

Extend the existing snapshot-bound metrics_{pr}.json authority with a native JSON
run_overengineering_audits boolean computed in Python from exact unified-diff
churn:

closes #4421

(added_lines + removed_lines) > 2000

The change adds packless, read-only reachability and abstraction-surface auditors,
keeps their dispatch independent from standard and deletion-regression review,
and carries accepted evidence through deterministic aggregation, publication, and
the resolve-review handoff.

Verification

• task test-check: 34,257 passed, 630 skipped, 27 expected xfails
• pre-commit run --all-files: all hooks passed
• independent implementation audit: verified GO with 32/32 requirements covered

Implementation Plan

Plan file: /home/talon/projects/generic_automation_mcp/.autoskillit/temp/make-plan/add_deterministically_gated_overengineering_auditors_plan_2026-07-29_202704.md

🤖 Generated with Claude Code via AutoSkillit

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested

[Trecek]

AutoSkillit review: blocking findings detected. See the 11 inline comments; changes are required.

Non-blocking adversarial observations:

• src/autoskillit/smoke_utils/_review.py:L78 — reachability: the new non-bytes stderr fallback is unreachable because the sole runner fixes capture_output=True and text=False; supported callers cannot replace the nested runner. Removing the fallback preserves reachable behavior and error context.
• src/autoskillit/smoke_utils/_experimental_review.py:L706 — abstraction surface: effect_artifacts duplicates documents already held by artifacts; the only production publisher reads artifact_order and artifacts, while direct effect_artifacts consumers are tests for the new projection. Removing the duplicate leaves ordering, rollback, concurrency, validation, and supported callers unchanged.

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested (self-authored PR; GitHub requires COMMENT event)

[Trecek]

AutoSkillit review: changes requested. One critical and ten warning findings were posted inline. Three informational observations are retained in the audit summary. Because this PR is self-authored, GitHub requires a COMMENT review event rather than REQUEST_CHANGES.

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested

[Trecek]

Test

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested (batch 3/4)

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested (batch 4/4)

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested (retry batch 1/4)

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested (retry batch 2/4)

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested (retry batch 3/4)

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested (retry batch 4/4)

[Trecek]

AutoSkillit PR Review — Verdict: changes_requested

74 findings after dedup; 2 critical, 52 warning, 20 info.
68 inline comments posted across 7 reviews (info/warning/critical).
6 findings could not be posted inline (lines outside diff hunks).
2 findings suppressed as duplicates of prior resolved threads.

Outside Diff Range

These findings target lines not in the diff and could not be posted as inline comments:

src/autoskillit/recipes/implementation.yaml

• L1082 [info/cohesion]: The same clear_review_annotation_context step block (callable, capture set, on_success/on_failure, note) is duplicated f

src/autoskillit/skills_extended/review-pr/SKILL.md

• L165 [warning/cohesion]: The hunk-ranges artifact is named inconsistently across layers: the on-disk file is ranges_{pr}.json (per _review.py L46

src/autoskillit/smoke_utils/_experimental_review.py

• L2630 [warning/cohesion]: EXPERIMENTAL_REVIEW_AUDITORS, _EXPERIMENTAL_DIMENSIONS, and the closed-set validator constants duplicate agent definitio
• L3025 [info/slop]: assert isinstance(candidate, dict) is dead code: _candidate_validation_error already verified the candidate is a dict an

src/autoskillit/smoke_utils/_review.py

• L3681 [warning/defense]: assert refs_before is not None after _read_pr_refs(required=True) is invoked. Under python -O the assertion disapp
• L3674 [warning/defense]: _required_scalar decodes git output with errors="strict". If a git rev-parse result contains non-UTF8 bytes, the Runti