daily-fix: implementer pin-sweep + lint parity

Overview / Motivation

Auto-filed by the /daily 2026-07-25 problem sweep. The implementer's self-declared verification scope and the gate's actual scope disagree in three separate ways, and the reviewer or the gate paid for it in 5 of the day's sessions.

Goal

Emit the pin-sweep hit list mechanically (selector-computed, with sweep_scope declared by the tool rather than by the implementer), always include the repo-wide invariant tests in the implementer's pre-commit local union when the diff touches any scripts/*.py or src/** file, and run the repo ruff-policy pin rather than a default ruff check in the implementer's lint leg.

Workflow gap

1. Pin-sweep hit lists under-enumerated in 4 of 6 sessions. Every code-reviewer re-ran its own repo-wide/tree-wide pin sweep and found files the implementer's declared hit list had omitted: 8 files (#1667 @ 08:12Z), 2 (#1669 @ 10:24Z), 8 (#1670 @ 09:57Z), 5 (#1671 @ 09:54Z). All were run green by the reviewer, so nothing shipped broken — but each cost the reviewer an extra sweep + test run and produced a recurring non-blocking Minor. #1667 verbatim: "gate-scope sweep omitted 8 generic-token hit files — I ran all 8 green"; #1670: "pin-sweep hit-list under-enumeration of 8 basename-fragment test files — all 271 run by me and green". The implementer's declared list and the reviewer's independent sweep are two different computations of the same thing; only one is mechanical.
2. The local union misses repo-wide invariants. #1681 (832cccf2) round 1 introduced PROJECT_ROOT / "tasks" at scripts/autonomous_session_watch.py:8220 — a direct violation of the always-on canonical-resolver invariant pinned by tests/test_no_direct_task_path_construction.py. The implementer reported a "4,208-test local union green" and the code-reviewer PASSed with 0 Critical / 0 Major; neither ran that invariant test. The Step 9c gate found it after a 27-min run: E 1 file(s) violate the canonical-resolver rule. … Matches: - scripts/autonomous_session_watch.py:8220. Cost a round-2 implementer spawn plus a second 29-min gate run (~40 min). The local union is derived from touched-file mapping, which does not pull in repo-wide invariant tests that ANY scripts/*.py edit can break.
3. Local ruff leg is narrower than the gate's ruff pin. #1672 (188282d2) @ 11:33:39Z — the implementer reported ruff check + format clean on its main commit, but the gate-scope union's tests/test_ruff_policy.py full-ruleset pin caught a UP033 violation, forcing a second commit (cfb4a2a297) purely to swap a decorator to functools.cache. The "ruff clean" acceptance criterion is therefore not the criterion the gate enforces.

• Confidence (emitter): high on all three (4/6, 1, and 1 same-day occurrences, each quoted from the sessions' own reviewer findings or gate output).
• verified-at-filing: the three always-on invariant tests exist and are named in CLAUDE.md § Common Commands as the workflow-invariant family — ls tests/test_no_direct_task_path_construction.py tests/test_no_pod_side_task_py_shellout.py tests/test_no_dollar_budget_caps.py tests/test_ruff_policy.py → all four present. The #1681 violation text, the four under-enumeration counts, and the UP033 fix-up commit are quoted from the sessions' own tool output / reviewer findings, not from recall; git rev-parse --verify --quiet 'cfb4a2a297^{commit}' resolves. Landed-fix history check git log --oneline --since='7 days ago' -- .claude/agents/implementer.md .claude/agents/experiment-implementer.md scripts/select_step9c_tests.py → the wave touched all three (#1682 841304c2d0, #1688 ab45d65777, #1673 37f6f6b1b4, #1651, #1649, #1646, #1645); none makes the sweep mechanical, adds the invariant family to the local union, or changes the lint leg's ruleset. (2026-07-25)

Proposed change (refine in planning)

+ (1) selector emits the pin-sweep hit list: the implementer RUNS the tool and pastes
+     its output (with the tool-declared sweep_scope) instead of hand-enumerating, so
+     the implementer's list and the reviewer's independent sweep are one computation.
+ (2) local union: when the diff touches any scripts/*.py or src/** file, always
+     include tests/test_no_direct_task_path_construction.py,
+     tests/test_no_pod_side_task_py_shellout.py, tests/test_no_dollar_budget_caps.py.
+ (3) lint leg: run tests/test_ruff_policy.py (or ruff with the pin's full ruleset)
+     rather than a bare `ruff check`, so "lint clean" means what the gate means.

(2) is cheap — those three tests are fast static scans, not suite runs. (1) is the structural one: if the selector already computes the hit list for the reviewer, the implementer should call the same entrypoint.

Scope / surfaces

• .claude/agents/implementer.md and .claude/agents/experiment-implementer.md (verification/smoke-run legs — both, or the shared text they point at).
• scripts/select_step9c_tests.py if (1) needs a new output mode rather than reusing an existing one — check first; the reviewer already runs a sweep, so the computation likely exists and only needs an implementer-facing entrypoint.
• Both agent specs are large: the agent-spec size ratchet (workflow_lint --check-agent-spec-size, WARN >28KB / FAIL >40KB) applies — prefer editing shared pointer text over duplicating three bullets in two files.

Constraints / invariants

• (2) must not balloon the local union into a full-suite run — name the three static invariant tests explicitly, do not widen to "all workflow-invariant tests".
• (1) must keep the implementer's declaration auditable: the reviewer's independent sweep stays as the check; this makes the two agree, it does not remove the check.
• scripts/workflow_lint.py --check-references / --check-asks pass; ruff passes (including tests/test_ruff_policy.py, which is the point of item 3).
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: .claude/agents/implementer.md
• fingerprint: 5095937ee216
• Source: /daily 2026-07-25 transcript sweep, sessions 203baf55 (#1667), 7457e1a3 (#1669), ad35514c (#1670), 25e73c77 (#1671), 832cccf2 (#1681), 188282d2 (#1672).