Tasks: Decide the fate of out-of-matrix test directories
Mission: decide-out-of-matrix-test-dirs-01M2TKC7 · Issue: #4374 · Branch: chore/decide-out-of-matrix-test-dirs Promotion policy: BOUNDED (operator-confirmed) — decide all 128 entries; promote only existing-row pop-(a) trees + at most the security-critical auth new-row via scrub; defer the rest as recorded decisions.
Why the WP shape is serial-ish
Most of this mission edits a small set of shared config files (.github/ci-module-registry.yml, .github/ci-shard-timings.json, tests/release/ci_retirement_scrub.json). Two WPs cannot both own the same file, so the registry/timings/scrub edits are consolidated into one execution WP (WP03), fed by an analysis WP (WP01), with two disjoint-file code WPs for lifted invariants (WP02) and permanent evidence guards (WP04).
Subtask Index
| ID | Description | WP | Parallel |
|---|---|---|---|
| T001 | Enumerate all 128 targeted dirs; classify recursive descendants (nested-family pre-check PC1) | WP01 | |
| T002 | Resolve owning row per pop-(a) tree (ownership tie-break PC2); flag rootless + promotion-blocked | WP01 | |
| T003 | Coverage-diff evidence for every disposition-3 candidate (demoted ⊆ in-matrix mirror) | WP01 | [P] |
| T004 | Marker-presence grep per pop-(b) tree; classify honest marker home vs #2979 gap | WP01 | [P] |
| T005 | Write the disposition ledger (dir → disposition → owner → rationale → evidence pointer) | WP01 | |
| T006 | Identify the lift-invariant set; specify each invariant + its always-on test | WP01 | |
| T007 | Author tests/architectural/test_lifted_*.py for each disposition-2 invariant | WP02 | |
| T008 | Sharpen the pop-(a) 60-dir mega-group into decision-grouped reasons naming in-matrix targets | WP03 | |
| T009 | Promote the capped existing-row pop-(a) set (parent-only, de-record descendants, non-nesting) | WP03 | |
| T010 | Capture timings for every promoted/expanded module; assert provenance parity; size shards (LPT) | WP03 | |
| T011 | Mint the auth row via scrub+registry verbatim edit; capture; size shards | WP03 | |
| T012 | Sharpen the pop-(b) 68-dir mega-group; reconcile integration with special-tier; nightly split if too-slow | WP03 | |
| T013 | Fold #4426 (decide tests/performance, tests/specify_cli/live_work); keep exclusions untouched | WP03 | |
| T014 | Run the full acceptance-checks contract; guard green; injected-red proof via gate_selection.py | WP03 | |
| T015 | Add permanent evidence guard: provenance parity + freshness (module_capture_provenance) | WP04 | |
| T016 | Add permanent marker-presence honesty guard for out-of-matrix marker-named reasons | WP04 | |
| T017 | Permanent guard: no out-of-matrix reason uses a deferral phrase / must name a covering surface | WP04 | |
| T018 | Permanent guard: each disposition-3 reason resolves to a committed coverage-diff JSON (subset) | WP04 | |
| T019 | Permanent guard: gate_selection selects the owning module for each expanded test_dir (per-PR proof) | WP04 |
Work Packages
WP01 — Disposition analysis & ledger (planning_artifact)
plan rather than re-deciding per tree. Security-first ordering (FR-008).
rationale + (where disposition-3) a coverage-diff pointer, and (where pop-(b) marker) a marker-grep result.
- Goal: Produce the authoritative per-dir disposition ledger with evidence, so WP02–WP04 execute a decided
- Priority: P1 (MVP brain). Independent test: ledger covers all 128 dirs, each with a disposition +
- Subtasks: T001, T002, T003, T004, T005, T006
- Depends on: none. Est.: ~420 lines.
WP02 — Lift repo-invariants into the always-on battery (code_change)
one drift-prone invariant (the #4715 pattern), so the tree can stay out-of-matrix with an honest reason.
and fails when the invariant is violated.
- Goal: For each disposition-2 tree, add an always-on
tests/architectural/test_lifted_<x>.pyasserting its - Priority: P2. Independent test: each new test is
pytest.mark.architectural, passes on green code, - Subtasks: T007
- Depends on: WP01 (ledger names the lift set). Runs PARALLEL to WP03. Est.: ~260 lines.
WP03 — Config-surface dispositions: sharpen + promote + auth-row + fold (code_change)
promote the capped set with measured evidence, mint the auth row via scrub, reconcile integration with the special tier (nightly split if too-slow), fold #4426. Guard stays green throughout.
reasons among targeted dirs; every shard_count change has fresh parity-checked evidence; injected-red proof passes through gate_selection.py.
- Goal: Execute every registry/timings/scrub decision: sharpen reasons (in-matrix-target + marker-checked),
- Priority: P1 (the execution). Independent test:
test_module_shard_registry.pygreen; zero snapshot - Subtasks: T008, T009, T010, T011, T012, T013, T014
- Depends on: WP01 only (parallel to WP02). One commit per subtask; guard green at each. Est.: ~560 lines (heaviest; the coupled-config core).
WP04 — Permanent evidence & marker-honesty guards + acceptance dossier (code_change)
promotion evidence parity/freshness and marker-presence honesty. Record the mission-acceptance summary.
violation.
- Goal: Lift the mission's load-bearing acceptance checks into always-on guards so future drift is caught:
- Priority: P2. Independent test: new guards pass post-WP03, and fail on a synthetic parity/marker
- Subtasks: T015, T016, T017, T018, T019
- Depends on: WP02 AND WP03. Est.: ~360 lines.
Dependencies
WP01 ─┬─► WP02 ─┐
└─► WP03 ─┴─► WP04WP02 (lifts, owns tests/architectural/test_lifted_*.py) runs in PARALLEL with WP03 (owns the config files) — disjoint ownership, and the lifted-test NAMES are fixed in the WP01 ledger, so WP03's reasons reference decided strings, not WP02's code. WP04 asserts the post-WP03 + post-WP02 state, so it depends on both. WP01→{WP02∥WP03}→WP04.
MVP
WP01 (the ledger) + WP03 (execution) are the mission MVP — they alone satisfy #4374's acceptance (every entry a decision, guard green). WP02/WP04 harden it (lifts + permanent guards).