daily-held: decide ruff-debt burn-down (2149 errors)

Overview / Motivation

Filed by the /daily 2026-07-02 backfill problem sweep (route 3 — genuine judgment call, needs Thomas). Held under the carve-out: bulk scope/priority decision with large blast radius (NOT auto-dispatchable).

The decision needed

main carries ~2,149 pre-existing repo-wide ruff errors (2,072 on 07-02, growing). On 2026-07-02 alone, at least 7 sessions each burned a verification round proving their diff didn't cause the red lint. Options:

1. One-time burn-down task — ruff check --fix + manual triage of the rest, as a dedicated kind: infra task. Cost: a huge diff that will conflict with ~15 live worktree branches; needs a quiet window.
2. Baseline-and-freeze — accept the debt, rely on the Step 9c known-red baseline ledger (filed separately) so sessions stop paying the re-derivation tax; ratchet: no NEW errors allowed.
3. Do nothing — status quo; every code session keeps paying the pre-existing-ness triage round.

Suggested: option 2 now (cheap, no conflicts), option 1 in a quiet window after the current experiment wave lands.

Decision (2026-07-17, task #1023 — autonomous under the 2026-07-17 PM greenlight)

• Option 2 (baseline-and-freeze): RATIFIED as standing policy. Mechanism = #1022's scripts/step9c_baseline.py ledger + live ratchet (landed 2026-07-17, before this task ran; the "filed separately" task in option 2's text). The Step 9c lint compare computes root-vs-worktree counts LIVE, so the ratchet survives baseline drops; only an INCREASE fails.
• Option 1 (bulk burn-down): DEFERRED + DESCOPED. Only 73/2,226 errors were safe-auto-fixable; 86% of the debt sat in frozen per-issue experiment scripts. After the config scoping below, the burn-down target is ~33 errors + 17 format-dirty files (~19 safe auto-fixes). Filed as #1486 (on_hold; revive trigger: <3 live issue sessions per spawn_session.py list, or 2026-08-01, whichever first).
• Option 3 (do nothing): REJECTED.
• Executed now: pyproject ruff scoping (config-only; branch commit dff56be0bc on issue-1023, squash-merged to main at Step 10d — see the epm:merged marker for the main SHA): extend-exclude gains eval_results + figures; [tool.ruff.lint.per-file-ignores] turns off style/noise rules on scripts/*, experiments/*, eps/experiments/*. Repo-wide count 2,226 → 33 (measured pre-merge in the worktree AND by CLI simulation at the repo root; format count unchanged at 17). Real-bug rules stay ON everywhere (F undefined-names/imports, E9 syntax, core bugbear B); src/ + tests/ keep the full ruleset; 35 live workflow helpers pinned full-ruleset- clean by tests/test_ruff_policy.py.
• E7 substitution (narrower than planned): the wholesale E7 ignore was replaced by specific E731 + E741 (the only E7xx codes in the frozen-path histogram) — E722 bare-except stays enforced at zero residual cost. Cost of the narrowing: future frozen scripts' other-E7xx hits become visible count growth (acceptable; they are real style regressions worth seeing).
• Accepted real-bug-adjacent losses on scoped paths (named deliberately): B905 (zip-strict), B023 (loop-var binding), F841 (unused variable) are off on scripts/* / experiments/* / eps/experiments/*. The code-review ensemble remains the gate on every new diff there.
• Honest justification for the scoping (not a "write-once" claim — 389/668 frozen-pattern scripts were re-touched in ≥2 commits over 30 days): (a) style enforcement was vacuous on the debt-accretion channel — experiment-branch merges never run ruff, so the rules never gated where the debt grew; (b) the Step 9c touched-file absolute-clean rule made touching one line of a frozen script demand fixing all its style errors — friction with no bug-finding value.
• Bulk --add-noqa: REJECTED (~2k-line diff across hundreds of files; noqa debt compounds — 776 files already carry noqa whose RUF100 interaction added +3,278 phantom errors in simulation; RUF100 is kept in every ignore list to guard exactly that explosion).
• Growth-channel merge-gate (ruff at experiment merges): NOT filed. The scoping removes new frozen scripts' style noise from the count. Re-check trigger: if the visible count grows >+5/week sustained over 2 weeks (style noise via src/ counts too, not just real bugs), file the workflow-fix candidate for a merge-time ruff ratchet then. Escalation observer: the watcher's gate-push Telegram channel + /daily problem sweep.
• Step 9c transition note (R1): a worktree cut BEFORE this change lints with its own old pyproject, so its Step 9c lint compare reads wt_ruff_count >> base_ruff_count with touched files clean — the main-side-cleanup false-FAIL step9c_baseline.py::lint_verdict pre-registers as one-glance diagnosable (~8 active infra-kind tasks in scope at decision time; experiments never run ruff). Recovery: rebase the branch onto origin/main and re-run. Revert recipe if the wave misbehaves (>48h multiple sessions blocked): git revert <merge-sha> — one hunk, direction-safe (a post-change worktree at 33 still passes against a reverted 2,226 root).

Provenance

• source: /daily 2026-07-02 backfill problem sweep (route 3, needs-human)
• carve-out: scope/priority judgment with large blast radius
• greenlight: PM directive 2026-07-17 (epm:progress v1) — autonomous background execution; decision executed via plan v2 (planner → fact-checker → 3-lens critic ensemble APPROVE ×3 + consistency PASS; Codex twins sentinel-skipped #1204)