[monkytheluffy]

harden-repo skill applied to ppgrid (Python): size ratchet (non-blank line budgets, no-growth) + top-level-unit export gate, wired into CI + pre-commit hook. Append-only: 5 files, 159 insertions, 0 deletions (ci.yml, AGENTS.md, pyproject.toml, scripts/size_ratchet.py, size-ratchet.json).

Adversarial review: PASS (3537407, 37m) — all 10 checklist items verified by independent re-run:

• both failure modes proven live: NEW BLOAT (900-line scratch → FAIL rc=1) and GROWTH (+100 lines to ratcheted file → FAIL rc=1); export gate fires on one added top-level def
• pre-commit hook proven both directions (temp commit blocked; clean commit passed full chain)
• ratchet seed byte-exact: reviewer's independent non-blank counts match the JSON exactly; --update regenerates byte-identical
• test parity with base under equal conditions: 623 passed / 2 skipped (worker's 609/16 was a venv extras artifact, not a commit effect)
• ruff check + format clean; CI job mirrors existing legs (fetch-depth 0, uv 3.13, --all-extras --locked)

Findings (ranked):

• LOW (acceptable): pyproject.toml per-file-ignores entry is load-bearing - with base config (select = ["ALL"]) the new script fails ruff 7 errors (S404/S601/S602/T201); the entry follows the table's existing convention
• NIT: scripts/size_ratchet.py:66 - nonexistent --base ref dies with raw CalledProcessError traceback instead of a clean gate message (CI fetch-depth 0 makes the ref always present; cosmetic)
• NIT: scripts/size_ratchet.py:101 - ast.parse on a non-Python base version of a >2000-line file raises uncaught SyntaxError (not reachable in this repo's CI today)

HUMAN-GATED (ppgrid): Andryo merges.

[monkytheluffy]

Adversarial review — monky (orchestrator lane, 2026-10-20)

VERDICT: PASS — no blockers. 5 files, +159/-0, pure gate, zero app-code changes.

Verified by execution (scratch worktree @ c0de4a5):

• Gate on branch: SIZE GATE: OK (4 ratcheted) rc=0, with and without --base main
• GROWTH mode: +100 lines to pipeline.py → GROWTH 6894 > budget 6794 rc=1
• NEW BLOAT mode: new 850-line file → rc=1
• EXPORT GATE mode: committed top-level def on pipeline.py (6794 lines) → EXPORT GATE: 114 top-level units (was 112) rc=1 (gate diffs base...HEAD, so committed state — correct for CI)
• Test-file x3: 2000-line tests/ file passes, 2500-line fails at 2400
• ruff check + ruff format --check (CI's exact --extra dev invocations): clean; the per-file-ignores names are valid
• CI wiring: standalone size job, no name clash with test/test-main/artifact, fetch-depth: 0 for the base ref, python 3.13 pin matches the other legs, push-to-main gets empty GITHUB_BASE_REF → ratchet-only as documented
• Baseline integrity: all 4 ratcheted values match current main
• AGENTS.md "pre-commit hook" claim: TRUE — the local hook runs size_ratchet.py --base HEAD (ratchet-only at commit)

Findings (ranked):

1. LOW — test-file budget is double-multiplied. The growth check is n > ratchet[rel] * mult, so a ratcheted test file gets TEST_MULT applied to its stored budget on top of the already-multiplied 800×3 flag. A test file ratcheted at 2500 lines would only fail at 7500 — 3x looser than the AGENTS.md reading ("ratcheted files may only shrink"; x3 is the flag budget). Latent today: no test file is ratcheted yet. One-token fix if the shrink-only reading is intended (gate on ratchet[rel], or store the multiplied budget at --update).
2. LOW — pre-commit hook is not reproducible from the repo. The size-gate line is appended to the local (untracked) .git/hooks/pre-commit; .pre-commit-config.yaml carries only the ruff hooks, so a fresh clone + pre-commit install loses the gate. Suggested fix: add it to the config as a local hook (entry: uv run python scripts/size_ratchet.py --base HEAD, pass_filenames: false) so pre-commit install reproduces it everywhere.
3. INFO — new >2000-line files fail the export gate (git show base:rel misses → prev=0 → any top-level unit fails). That is the intended force-a-split-from-day-one behavior; noting so nobody reads it as a false positive.
4. INFO — stale ratchet entries are never pruned. A file that shrinks below the flag keeps its budget entry forever (harmless; the JSON stays small).
5. INFO — AST parse is strict-UTF-8 while line counting is errors="replace". A non-UTF-8 .py crashes the gate with a traceback (still exit 1) instead of a gate FAIL line; the test leg catches syntax issues first in practice.

Findings 1-2 are worth a follow-up commit but do not block the merge: the gate as shipped is deterministic, strictly fails on growth, and is wired into CI + pre-commit on this box.