⚡ Do This First: Load Agent Profile

/ad-hoc-profile-load python-pedro; apply it and state what you applied. PYTHONPATH=$(pwd)/src .venv/bin/python -m pytest … (never a bare uv run).

Objective

Make the performance-marker guard, the CLI startup-budget guard, and the repo-root status guard discriminate. Read ../spec.md (US3/FR-003, US6/FR-006, US1/FR-001) and ../plan.md. RED-first non-vacuity demo per finding. C-003: test-infra only — zero src/ change (verified safe by the scout: no dynamic jsonschema import in src/).

Owned files (touch ONLY these)

tests/architectural/test_performance_marker_guard.py, tests/performance/test_cli_startup_budget_4409.py, tests/_support/repo_root_status_guard.py, tests/test_repo_root_status_guard.py, pytest.ini, .github/workflows/ci-nightly.yml.

Subtasks

T010 — #4210 additive real-tree walk + marker desc + env-scope invariant

In test_performance_marker_guard.py: the guard runs only on planted TEXT (_MISMARKED_FUNCTIONAL_PERFORMANCE_TEST etc.), never AST-walks the real tree. ADD _iter_performance_marked_test_sources() + a test that AST-parses each real tests/*/test_.py and calls the UNCHANGED pure find_functional_assertions_under_performance_marker(source) per file — do NOT change that function's signature (test_timing_coverage_invariant.py:51/525 imports and calls it). Bound the walk: rglob("test_*.py"), try/except SyntaxError, explicit exemption list; if it exceeds sub-second, drop the module's fast marker. Correct the stale pytest.ini marker desc (:58 "not selected by any live CI job" — the nightly -m performance job runs it). The "e2e double-run" is a GHOST (already removed by #4865): instead add a REAL invariant asserting SPEC_KITTY_RUN_PERFORMANCE is set on the performance nightly job ONLY (not leaked to any other job). Any ci-nightly.yml edit must keep sibling assertions green (test_marker_job_completeness.py, this file's own workflow tests :217-384). RED-first demos: a planted real @pytest.mark.performance test with a functional assertion is caught; a planted SPEC_KITTY_RUN_PERFORMANCE on a non-performance job is caught.

T011 — #4536 startup-budget floor + dynamic-import + docstring

In tests/performance/test_cli_startup_budget_4409.py: test_jsonschema_stays_out_of_module_scope builds offending from (REPO_ROOT/"src").rglob("*.py") (:92-96) with NO min-count assertion → a mis-rooted/empty glob passes vacuously; add assert len(scanned) > <floor>. Extend _module_scope_jsonschema_imports (:55-83) to also flag importlib.import_module("jsonschema") / __import__("jsonschema") string args. Correct the docstring (:18-19/:88-91) — it walks all of src/, not "the CLI's import graph". RED-first demos: a mis-rooted/empty scan fails the floor; a planted dynamic import_module('jsonschema') (in a fixture string, NOT src/) is flagged.

T012 — #4036 window-scope the status-guard self-test + document straddle

In tests/test_repo_root_status_guard.py: the self-test assert not (REPO_ROOT / name).exists() (:163-164) is an absolute claim (flaky under -n auto, doesn't express "this test didn't leak"). Window-scope it to the guard's per-test window. In tests/_support/repo_root_status_guard.py: document the concurrent-worker/between-tests straddle in the "Boundary" block (:52-56). RED-first demo (must be REAL — the scout flagged this as the easiest to fake): prove a write that STRADDLES the per-test window (setup snapshot → teardown compare, :166-188) is now caught/correctly-attributed — not merely that a comment was added. The guard already has a genuine e2e firing proof (:142-164); this is a narrow discrimination tightening.

T013 — demos permanent

Each demo committed RED-first (guard vacuous on base) then GREEN; leave as the guards' permanent non-vacuity tests (docstring-pin issue #).

DoD

Each of #4210/#4536/#4036 has a demo RED on base / GREEN after (record evidence; #4036's demo proves a real straddle).
git diff --stat shows ZERO src/ files (C-003/NFR-003).
Targeted: PYTHONPATH=$(pwd)/src .venv/bin/python -m pytest tests/architectural/test_performance_marker_guard.py tests/architectural/test_marker_job_completeness.py tests/architectural/test_timing_coverage_invariant.py tests/performance/test_cli_startup_budget_4409.py tests/test_repo_root_status_guard.py — record counts (the timing-coverage sibling MUST stay green → proves the pure-fn signature preserved). ruff/ruff-format/mypy clean; complexity ≤15.
RED-first demo commit(s) before the fix. End messages with the Co-Authored-By/Claude-Session trailer.

Reviewer guidance

Verify: the pure perf fn signature is UNCHANGED (test_timing_coverage_invariant.py green); the #4210 tree walk is bounded/non-flaky; #4036's demo proves a real straddle (not comment-only); the e2e-ghost was re-scoped to the env-scope invariant (not a no-op); zero src/ files touched.

Spec Kitty

Kanban Board