Mission Specification: CI suite-map bind — marker→job authority + coherence invariants (Wave 0 full)
Mission Branch: tidy/ci-suite-map-2034 Created: 2026-07-04 Status: Draft (rev 4 — plan-review squad folds: FR-010 de-vacuoused + run_all mechanism (HiC ruling), FR-011 decision table + C-005 aggregator scope, FR-012 narrowed to the real gap, FR-013 ready_for_review trigger (HiC ruling), reverse-containment, #1933 reconciliation) Input: Roadmap Wave 0 (docs/plans/degod-unshim-roadmap.md:40) under epic #1868 WS5; executes #2297 + #2296; closes the #2034 marker gap and #2283 factor (a).
Adjudicated Decisions (binding — do not re-litigate)
1. CI-VALIDATED invariants, NOT a workflow generator. #2297's body says "GENERATES", but the governing WS5 acceptance criterion reads "generated or CI-validated; divergence fails an architectural test" — and the codebase's entire CI-enforcement precedent is parse-and-assert (_gate_coverage.py, test_ci_quality_path_filters.py, test_marker_registry_single_source.py). A generator over the hand-tuned 2,915-line ci-quality.yml would mint a third source of truth (split-brain risk) with no precedent. The "machine-readable suite map" deliverable is the parsed gate model (_gate_coverage.load_gates()) plus the invariant suite bound over it. (alphonso Q1; rejected alternative recorded in research.) 2. The #2034 census is STALE — this is a binding-layer mission, not a rescue. Live census (debbie, 2026-07-04 @595c50ff8): PR #2294 already bound 365/369 orphan files; PR #2047 already single-sourced the marker registry (pytest.ini authoritative, guarded); ALL 9 named hidden failures now PASS (kwarg drifts fixed by degod waves, snapshots by PR #2319, SEO page deleted, home-literals cleaned); the parallel-collection nondeterminism is fixed (sorted(...) + docstring). True residue: 17 zero-gate tests / 4 files live (19/5 at census, glossary lock fixed pre-spec; all green) + ~273 marker-only tests mostly covered by path-only gates. Do not spec against the issue body's numbers. 3. Residual job, NOT a -m "unit or contract" job and NOT shard-widening. A dedicated unit or contract job would run 11,088 tests (~39% of suite, ~10.8k double-run). The gate is a small always-on job selecting only the residual expression ((unit or contract) and not (<every routed runnable marker>)) — ~252 pure-logic tests at spec time (renata re-derived; re-derive at implement per NFR-004), seconds. (alphonso Q3.) 4. Marker-completeness is name-level; the existing orphan ratchet stays set-level. Two complementary invariants: the new one asserts every registered marker NAME is positively referenced by ≥1 gate or CI-invisible-with-reason; test_gate_coverage.py keeps asserting every test's marker SET reaches ≥1 gate. No duplication of combination-reachability. (alphonso Q2.) 5. Fix-all-before-gate; a THREE-STATE marker model (renata HIGH-2 fold — honest sizing). Live introspection: only 8 of 37 registered markers are positively referenced in any gate -m expression (fast, integration, git_repo, slow, architectural, timing, quarantine, windows_ci); after FR-002 routes unit/contract, ~27 markers would red under a naive two-state invariant — but ~15 of those (e2e, doctrine, agent, upgrade, distribution, orchestrator_*, regression, adversarial…) are CI-visible via PATH gates, not invisible. Forcing them into CI_INVISIBLE would mislabel running tests as invisible (the C-003 dumping-ground failure). The invariant therefore models THREE states: ROUTED-BY-MARKER (positive -m token), ROUTED-BY-PATH (every collected test carrying the marker is reachable via ≥1 path gate — verified against the orphan model, not asserted by hand), and CI_INVISIBLE (reasoned allowlist for genuinely-unrun markers: platform_darwin, live_adapter, exploratory, flaky, non_sandbox, timeout, asyncio, stress…). Exact membership derived and recorded at implement time, shrink-preferred. unit and contract are INELIGIBLE for both CI_INVISIBLE and ROUTED-BY-PATH — the invariant hard-asserts they are ROUTED-BY-MARKER (renata MEDIUM-3: otherwise the allowlist could absorb them and defeat the mission). No #2317-style quarantine lane — the residual set is pure-logic and currently green. (alphonso Q5 + renata folds.) 6. Out of scope with tracked homes: #2283 factor (b) local venv skew → CT7/#2077 or a #2283 sub-item; factor (c) producer-conformance sweep → CT7/#2077; quarantine triage → #2295/#2309 (the new gate excludes quarantine by design and cannot re-surface them); mypy CI-scope expansion (WS5 sub-AC) → #1868 follow-up (needs an advisory-vs-enforced decision this mission doesn't own); WS5 path-topology tail (#2333) → FOLDED IN-MISSION (operator ruling 2026-07-04: the remainder of Wave 0 is thin; deferring hurts) — AC-2a src-side filter coverage, AC-6 skipped-suite fail-closed + step-summary visibility, AC-8 guard self-mapping, AC-1's --ignore mirror coherence are FR-010..FR-012; #2333 is closed by this mission's PR; marker-registry dedup → DONE (PR #2047), depend on it. 7. HiC rulings on the path-topology mechanism (operator, 2026-07-04, after the plan-review squad): (a) ready_for_review IS added to ci-quality.yml's on: pull_request types: (FR-013) — without it a draft→ready PR never re-runs its draft-skipped suites (live merge-safety hole); with it, GitHub's native draft-merge block + the re-trigger make FR-011's draft exemption safe. (b) The FR-010 catch-all forces run_all semantics (every suite) via the existing :84-103 OR-seam — a LOUD ALARM, not steady state: closeout adds named groups for hot unmatched dirs and records a shrink-trend obligation, reconciling #1933's keep-PR-CI-targeted intent (unmatched-hits trend to zero; mapped PRs stay targeted). Routing unmatched changes only to the core-misc jobs was REJECTED (they --ignore per-module dirs — wrong suites; paula). The static-only alternative was REJECTED (leaves the gap-window unrun). 8. Path-topology authority rule (paula Q5): path→suite topology has exactly TWO source authorities — the dorny filter block (path→group) and the job if: gates (group→job). Every derived surface this mission adds (the catch-all's group-reference OR-list, the aggregator's job→group table, the invariants) must be ASSERTED-AGAINST the parsed sources, never hand-maintained in parallel. A free-standing literal copy of either mapping is a REJECT (it would create the split-brain this mission exists to close).
Squad-Verified Census (authoritative over issue bodies)
- Gate inventory: every
ci-quality.ymljob selects by marker (fast/integration/git_repo/architectural/slow/timing/quarantine+ e2e-by-path);mission-loader-coverageand the focused lint steps are path-only;slow-testsruns on push only (PR-invisible, push-visible — documented, not a defect this mission fixes); the nightly schedule collapses path filters but never marker filters; no gate selects-m unitor-m contract(confirmed live). - Orphan population: authoritative path+marker model reported 19 orphan tests / 5 files at census; the 5th file (Wave 2's glossary lock) was fixed pre-spec (marked
fast), returning the live floor to 17 tests / 4 files — matching the committed_gate_coverage_baseline.json(orphan_test_count: 17). Re-derive at implement (NFR-004). - Residual population: ~273 nodes / 29 files carry
unit/contract-class markers with no runnable routed marker (marker-only view); the residual EXPRESSION collects ~252 (renata re-derived); most are covered by path-only gates today; the residual job (FR-002) makes the whole class marker-reachable regardless. - Enforcement substrate:
_gate_coverage.pyparses all 4 suite-running workflows intoGate(paths, ignores, marker_expr)and evaluates real collected tests via pytest's ownExpression;test_gate_coverage.pyis the file-level orphan ratchet;test_ci_quality_path_filters.py+test_ci_architectural_gate_coverage.pykeep dorny filters honest;test_marker_registry_single_source.pypins pytest.ini as the registry. - WS5 live-gap status (epic census at
aec70048vs this branch): dark suitetests/runtime/→ already routed (fast-tests-next runs it) — verify-and-record, no work; phantomneeds:(integration-tests-charter.result) → renata verified LIKELY ALREADY FIXED* (declared at ~:2938 AND read at ~:2988) — re-verify at plan time and record verified-already-fixed if so;quality-gateaggregator excludesmission-loader-coverage→ in FR-004's surface; vacuoussrc/mission_runtime/diff-coverage entry → in FR-004's surface (cheap same-parse invariant).
User Scenarios & Testing (mandatory)
User Story 1 - A test authored with the documented default marker cannot be CI-invisible (Priority: P1)
A contributor writes a pure-logic test and marks it unit (the authoring default pytest.ini documents) or contract. Today that marker routes to no job — the test runs only if its file happens to carry a routed sibling marker or sit under a path-only gate. After this mission, unit/contract are positively selected by a real gate, and a registered marker that loses (or never gains) a selecting job fails an architectural test at PR time.
Why this priority: this is the #2034/#2283(a) root cause and the roadmap's Wave-0 rationale — every later degod wave's pure-core tests must actually run pre-merge.
Independent Test: register a synthetic marker in pytest.ini without routing it → the marker-completeness test REDS naming it; add it to CI_INVISIBLE with a reason → green; remove unit from the residual job's expression → REDS.
Acceptance Scenarios:
1. Given the merged branch, When the marker-completeness arch test runs, Then every marker in pytest.ini's markers = block is either positively referenced (non-negated token) in ≥1 gate's -m expression across the 4 suite-running workflows, or present in CI_INVISIBLE with a non-empty reason — and the test is sub-second (pure expression walk, no collection). 2. Given a PR that registers a new marker with no job and no allowlist entry, When CI runs, Then the architectural shard fails naming the marker (fail-closed on NEW). 3. Given a PR that edits a gate's -m expression such that a routed marker loses its only positive reference, When CI runs, Then the same test fails (fail-closed on DE-ROUTING; same mechanism, no separate machinery).
User Story 2 - The unit/contract residual actually runs, cheaply (Priority: P1)
The ~252 tests carrying only authoring-taxonomy markers get a standing, always-on CI home that costs seconds — without double-running the ~10.8k unit tests that already ride fast/integration shards.
Why this priority: the gate is what makes US1's invariant satisfied structurally (the invariant demands a selecting job exist; this is that job).
Independent Test: pytest -m "<residual expression>" --collect-only -q collects ≈252 tests (re-derive) and the full selection passes locally in under a minute; the job appears in quality-gate's needs: and result loop.
Acceptance Scenarios:
1. Given the new unit-contract-residual job, When it runs on any PR (always-on or src/tests-triggered; not draft-gated), Then it selects the residual expression, excludes windows_ci and quarantine per convention, and is a BLOCKING input to quality-gate (declared in needs: AND read in the result loop — no phantom). 2. Given the residual set as of spec time, When the job first runs, Then it is green (debbie: all 17 live orphans + residual sample currently pass; gate-unmask discipline — any red found at implement time is fixed or re-marked before the gate lands, never quarantined-by-default). 3. Given a future test authored unit-only, When it lands, Then it is selected by this job (no path dependence).
User Story 3 - The workflow's duplicated topology knowledge cannot drift silently (Priority: P2)
The WS5 coherence invariants bind the relation subset this mission owns: every needs.<job>.result read in a run script is declared in that job's needs:; every dorny filter output is consumed by ≥1 job if:; every filter glob matches ≥1 tracked path; every diff-coverage critical-path entry is backed by a job emitting a matching --cov; every [ENFORCED] per-package gate is reachable from quality-gate. (The path-topology ACs — src-side filter coverage, skipped-suite fail-closed, guard self-mapping, --ignore mirrors — are IN-MISSION as US5/FR-010..FR-012 per the operator fold ruling; #2333 tracks them and closes with the PR.)
Why this priority: these are the WS5 AC(b)-(f) relations; the phantom-needs: class means a red suite can coexist with a green aggregator — the worst failure mode of the whole delivery topology.
Independent Test: each invariant has a fault-injection test proof at implement time (e.g., add a needs.<x>.result read without declaring <x> in a fixture copy → REDS).
Acceptance Scenarios:
1. Given the invariant suite, When it parses the live workflows, Then all live incoherences found at implement time have been FIXED in ci-quality.yml in the same mission (verified candidates from the epic census: the integration-tests-charter phantom read; mission-loader-coverage reachability from quality-gate; the vacuous src/mission_runtime/* diff-coverage entry) — and the invariants pass on the fixed workflow. 2. Given any future PR that reintroduces one of these incoherence classes, When CI runs, Then the architectural shard fails naming the relation violated. 3. Given the invariants, When a job is renamed or the YAML reordered without changing behavior, Then nothing reds (refactor-stable: keyed on parsed relations, never job names at line numbers or literal -m strings — C-006).
User Story 4 - The orphan baseline drains to zero (Priority: P2)
The 4 files frozen in _gate_coverage_baseline.json (#2296) get their ci-quality.yml path additions (workflow-scope-gated), the baseline regenerates to zero orphans, and the ratchet floor locks there.
Why this priority: the clean termination condition for the Wave-0 orphan work; small, but it's the difference between "ratchet with a tolerated floor" and "zero-orphan invariant".
Independent Test: python -m tests.architectural._gate_coverage --check (or its test wrapper) reports orphans: 0; the baseline file records 0; the ratchet still fails-closed on any new orphan file.
Acceptance Scenarios:
1. Given the two tests/coordination/ resolver tests and tests/paths/test_windows_migrate.py, When their dirs are added to the matching integration shard paths, Then they are selected by ≥1 gate and leave the baseline. 2. Given tests/_support/git_template/test_git_template.py (the issue flags "may be intentionally non-gated — decide"), When the mission adjudicates it, Then it is either routed like the others or explicitly recorded as intentionally non-gated with a rationale in the baseline/allowlist — no silent tolerated floor. 3. Given ci-windows.yml's trigger paths (WS5 AC), When the mission lands, Then the Windows job's path triggers cover src/** (or the recorded per-package rationale for exclusions) so Windows-sensitive src/mission_runtime changes trigger it.
User Story 5 - A src change can never silently skip its suites (Priority: P1, folded #2333)
Today a change confined to src/glossary (zero filter groups) or 43% of src/specify_cli children triggers the workflow but ZERO test jobs, and quality-gate passes because skipped-is-OK. After this mission: every src/ surface is either matched by a filter group that gates a real job, or caught by a fail-closed catch-all that forces the full-run path; a changed-and-mapped-but-skipped suite is visible (step-summary run/skipped table) and the aggregator treats an improperly-skipped mapped suite as a failure; the guard tests and workflow files are themselves mapped (a gate edit re-runs the guards); the two hand-mirrored catch-all --ignore lists cannot drift from the filter topology.
Why this priority: this is the delivery-topology half of the same disease — the marker invariant (US1) closes the taxonomy hole, this closes the path hole. Operator ruling: too load-bearing to defer.
Independent Test: touch a file in src/glossary on a probe branch → the catch-all (or a new group) triggers ≥1 test job; simulate a mapped-suite skip with changed paths in a fixture → the aggregator arm reds; remove a dir from a catch-all --ignore without a filter change in a fixture → the mirror invariant reds.
Acceptance Scenarios:
1. Given the filter block post-mission, When the src-coverage invariant runs, Then every top-level src/ package and every second-level src/specify_cli child is matched by ≥1 filter group whose output gates ≥1 test-running job, OR falls to the fail-closed catch-all group that forces the full-run path — zero silently-unrouted surfaces (AC-2a, fail-closed by construction rather than 75 hand-enumerated groups). 2. Given a PR whose changed paths map to a suite that did not run, When quality-gate evaluates, Then the run/skipped table appears in $GITHUB_STEP_SUMMARY and an improperly-skipped mapped suite fails the gate (AC-6; "improperly" = its filter output was true but the job did not run and was not superseded by the full-run path). 3. Given an edit to .github/workflows/ci-quality.yml or tests/architectural/, When the filters evaluate, Then ≥1 test-running job (incl. the architectural shard) triggers (AC-8 guard self-mapping). 4. Given the two catch-all --ignore lists, When the mirror invariant runs, Then** each list is exactly the set of test roots owned by dedicated shards (parsed relation, not literal duplication) — drift reds (AC-1 partial).
Edge Cases
- A marker positively referenced ONLY by a non-blocking gate (e.g.
quarantineinquarantine-visibility, which quality-gate ignores): the completeness invariant counts it as ROUTED (a job selects it) — blocking-ness is a separate, documented axis;quarantine's non-blocking-by-design nature is recorded in the allowlist reasons, not conflated with reachability. windows_ciis negated in every Linux gate but positively selected byci-windows.yml— the positive-token extractor must be negation-aware (pytestExpressionAST walk), ornot windows_ciwould falsely satisfy it.- The residual expression must be derived from the ROUTED-marker set, not hardcoded prose: when a new routed marker is added, the residual expression needs updating — the marker-completeness test should assert the residual job's expression stays consistent with the routed set (behavioral relation, not literal string pin).
- Draft-PR gating:
integration-tests-core-miscis draft/WIP-gated; the residual job must NOT be, or a draft PR could merge-queue with the residual unrun (verify quality-gate semantics for drafts at implement time). - The 4 suite-running workflows are the parse surface (
ci-quality,ci-windows,drift-detector,release); a NEW pytest-running workflow added later must enter the model —_gate_coverage's workflow list is itself an allowlist; assert it matches the live.github/workflows/*.ymlset that invokes pytest (glob + content probe), so a fifth workflow fails closed rather than silently escaping the model.
Requirements (mandatory)
Functional Requirements
| ID | Requirement | Verification | Status |
|---|---|---|---|
| FR-001 | A new architectural test (tests/architectural/test_marker_job_completeness.py or equivalent) asserts: every marker registered in pytest.ini is in exactly one of three states — (i) ROUTED-BY-MARKER: positively referenced (negation-aware token extraction via pytest's Expression AST) in ≥1 gate's -m expression across all suite-running workflows; (ii) ROUTED-BY-PATH: every collected test carrying the marker is reachable via ≥1 path gate, verified against the _gate_coverage orphan model (not hand-asserted); or (iii) CI_INVISIBLE: reasoned allowlist entry. Reverse containment holds: every CI_INVISIBLE entry (and every ROUTED-BY-PATH claim) must reference a marker REGISTERED in pytest.ini — a marker deleted from the registry but left in the ledger reds (renata: the dumping-ground failure, other direction). unit and contract are HARD-ASSERTED ROUTED-BY-MARKER — ineligible for states (ii) and (iii). Fails closed on new unrouted markers and on de-routing edits. State (i)/(iii) checks are collection-free; state (ii) may reuse the orphan model's cached collection. | Fault-injection: synthetic marker → red; allowlist → green; de-route unit → red EVEN IF allowlisted (ineligibility assertion). | Planned |
| FR-002 | A new small CI job (working name unit-contract-residual) in ci-quality.yml selects the residual expression — (unit or contract) and not (<routed runnable markers>) — excludes windows_ci/quarantine, runs on every PR (not draft-gated), and is a blocking input to quality-gate (declared in needs: and read in the result loop). | Job green on the ~252-test residual; collect-only count recorded; quality-gate wiring asserted by FR-003's invariants. | Planned |
| FR-003 | Workflow-coherence invariants (same parse model as FR-001, extending _gate_coverage): (a) every needs.<job>.result reference in a run script is declared in that job's needs:; (b) every dorny filter output is consumed by ≥1 job if:; (c) every filter glob matches ≥1 tracked path; (d) every job named in quality-gate's result loop is in its needs: and vice-versa for blocking jobs. Refactor-stable: parsed relations only. | Fault-injection per relation on fixture YAML; live workflows pass post-FR-004. | Planned |
| FR-004 | BOUNDED fix-set (renata MEDIUM-4): the named candidates — mission-loader-coverage unreachable from quality-gate; the vacuous src/mission_runtime/* diff-coverage entry (both re-verified at plan time) — plus any SAME-CLASS incoherences the FR-003/FR-005 invariants surface (needs-declaration, filter-consumption, glob-liveness, cov-backing). The integration-tests-charter phantom read is LIKELY ALREADY FIXED (renata: declared ~:2938, read ~:2988) — re-verify and record. Incoherence classes beyond the FR-named invariant set stay bounded; the path-topology classes are now owned by FR-010..FR-012 (folded #2333). | FR-003/FR-005 suites green on the fixed live workflows; each fix (and each verified-already-fixed) listed in the PR body. | Planned |
| FR-005 | Diff-coverage critical-path backing invariant: every entry in the diff-cover critical-path list is matched by ≥1 job emitting a --cov for that package (closes the vacuous-entry class). | Fault-injection + live pass after FR-004 fixes. | Planned |
| FR-006 | #2296 drain: the 3 route-able orphan files get ci-quality.yml path additions on their matching integration shards; tests/_support/git_template/test_git_template.py is adjudicated (route or record intentionally-non-gated with rationale); _gate_coverage_baseline.json regenerates to orphan_test_count: 0 (or the single recorded intentional exception) and the ratchet locks the new floor. | --check reports 0 orphans (± recorded exception); ratchet fault-injection still red on a new orphan. | Planned |
| FR-007 | ci-windows.yml trigger paths widened to src/** (or an explicit per-package exclusion list with rationale), so Windows-sensitive packages (src/mission_runtime) trigger the Windows job (WS5 AC). | Path filter includes src/**; recorded rationale for any exclusion. | Planned |
| FR-008 | Workflow-set completeness: the _gate_coverage model's workflow allowlist is asserted against the live set of pytest-invoking workflows in .github/workflows/, so a new suite-running workflow fails closed instead of escaping the model. | Fault-injection: fixture workflow invoking pytest not in the model → red. | Planned |
| FR-010 | Src-side filter coverage (AC-2a) with a NON-VACUOUS fail-closed catch-all: (a) mechanism — a trivial any_src: ['src/**'] probe group + a script step that computes unmatched = any_src AND NOT (union of all named group outputs) FROM THE FILTER'S OWN OUTPUTS (the negated mega-glob mirror is FORBIDDEN — it would hand-maintain a copy of all groups); unmatched threads into the existing run_all OR-seam (~:84-103) so every per-module job forces on (run_all semantics per HiC ruling 7b, no per-job if: surgery); (b) targeted named groups where a natural owner shard exists (src/glossary at minimum); (c) the invariant is the PARSED RELATION: the unmatched-computation's group-reference set ≡ the parsed filter-group set (a group added/removed without catch-all wiring reds) AND every group output gates ≥1 test-running job — NOT "every package matched-or-caught" (vacuous by construction once a catch-all exists; paula CRITICAL); (d) refactor-stability red-negative: reordering/renaming groups without changing the parsed relation stays green. | Fault-injection: fixture group not in the OR-list → red; fixture group reorder → stays green. Probe branches: unmatched src file → run_all path fires; well-mapped file → catch-all does NOT fire (fixture-asserted via the parsed model, plus one live probe recorded). | Planned |
| FR-011 | Skipped-suite fail-closed visibility (AC-6), scoped and non-fakeable: quality-gate emits a run/skipped table to $GITHUB_STEP_SUMMARY and FAILS per the decision table — FAIL iff filter_true AND job_skipped AND NOT full_run AND NOT draft_exempt (full_run = inputs.run_all OR catchall.unmatched; draft_exempt = job draft-gated AND PR is draft — safe per FR-013 + GitHub's native draft-merge block). SCOPE: blocking needs: jobs ONLY — quarantine-visibility stays non-blocking (C-005 extended: FR-011 must not re-surface the quarantine set into the blocking path; renata). The job→group mapping the arm consumes is ASSERTED ≡ the parsed job-if: gating by an FR-003-class invariant (Decision 8) — never a free-standing bash table. The decision logic is EXTRACTED into a testable script with fixture tests incl. the draft-context case and the improperly-skipped case. | Fixture tests on the extracted script (draft case, skipped-mapped case, full-run-supersede case); mapping-invariant fault-injection; live green post-fix; the table visible in the C-007 probe run (non-draft probe — core-misc is draft-gated). | Planned |
| FR-012 | Guard self-mapping (AC-8, narrowed to the REAL gap) + --ignore mirror coherence (AC-1 partial): ci-quality.yml + tests/architectural/** ALREADY self-map via core_misc (:175/:187 — record verified-already-fixed; renata/paula convergent); the live gap is the OTHER suite-running workflows (ci-windows.yml, drift-detector.yml, release.yml) which appear in NEITHER the outer on: paths: NOR any dorny group — the fix is TWO-LAYER: widen the outer on: pull_request/push paths: AND add a dorny group routing them to the architectural shard. The --ignore mirror invariant asserts each catch-all --ignore list ≡ the shard-owned test-root set as a PARSED relation with a refactor-stability red-negative (reorder/rename → stays green) — a literal-mirror implementation is a review REJECT (fault-injection alone cannot discriminate it; renata). | Fault-injection per relation; reorder red-negative; live green post-fix. | Planned |
| FR-013 | ready_for_review added to ci-quality.yml's on: pull_request types: (→ [opened, synchronize, reopened, ready_for_review]; HiC ruling 7a) so a draft→ready transition re-runs the draft-skipped suites; verified by a fixture assertion on the parsed trigger types + the C-007 probe (draft→ready flip re-triggers). | Parsed-trigger invariant + probe evidence. | Planned |
| FR-009 | Tracker closeout: #2297 closed by the PR; #2296 closed by the PR; #2034 closed by the PR (its residue is exactly FR-001/FR-002 — verify no remaining open obligation in its thread at closeout); #2283 gets a partial comment (factor (a) closed; (b)/(c) remain, pointed at CT7/#2077); #2333 closed by the PR (folded in-mission); #1933 gets the reconciliation comment (catch-all = loud alarm + shrink obligation; mapped PRs stay targeted); #1868 WS5 checklist comment updating the bound ACs; roadmap Wave-0 row struck. | Comments/links verifiable; issue-matrix terminal. | Planned |
Non-Functional Requirements
| ID | Requirement | Threshold | Status |
|---|---|---|---|
| NFR-001 | FR-001/FR-003/FR-005/FR-008 invariants are cheap enough for the architectural shard | combined < 5 s (pure parse + expression walk; no pytest collection) | Planned |
| NFR-002 | The residual job's wall-clock stays small | < 3 min end-to-end on CI (≈252 pure-logic tests at spec time; measure at implement time and record) | Planned |
| NFR-003 | No test double-run regression: the residual expression must not select tests already routed via fast/integration/etc. | duplicate-count from _gate_coverage --check does not increase vs the RE-DERIVED pre-mission value (committed baseline records 2,944; live census saw 3,291 — re-derive at implement, pin the honest number, shrink-only from there) | Planned |
| NFR-004 | Honest live counts (Wave-1/2 doctrine): every census number in the mission's artifacts (252, 17, 2,944…) is re-derived at implement time, never copied from this spec | re-derivation commands + outputs recorded in WP reports | Planned |
| NFR-006 | Catch-all breadth stays bounded: the unmatched signal fires ONLY when no named group matched (fixture-asserted via the parsed model); the closeout records the live unmatched-dir census + the shrink-trend obligation (#1933 reconciliation) | fixture assertion + census recorded in the PR body | Planned |
| NFR-005 | Every new/relocated test carries CI-selected markers (#2034's own lesson; proven for the new arch tests by FR-001 itself) | new tests selectable by ≥1 gate; verified per-file | Planned |
Constraints
| ID | Constraint | Rationale | Status |
|---|---|---|---|
| C-001 | No workflow generator; no third source-of-truth artifact. The parsed-model + invariants form is binding (Adjudicated Decision 1). | WS5 "or CI-validated" arm; precedent; split-brain risk. | Accepted |
| C-002 | Refactor-stable invariants only: parsed behavioral relations; never job names @ line numbers, literal -m strings, or YAML shape pins (testing-principles styleguide / C-006 doctrine). | The invariants outlive workflow refactors. | Accepted |
| C-003 | Shrink-only: CI_INVISIBLE and any baseline exception carry per-entry reasons; additions are loud (the completeness test names them), removals silent. | Exception ledgers must not become dumping grounds. | Accepted |
| C-004 | workflow OAuth scope required for ci-quality.yml/ci-windows.yml pushes — verify the push path early (WP01-adjacent), before deep work stacks behind it. | #2296 explicitly flags the scope gate. | Accepted |
| C-005 | Do not touch quarantined tests (#2295/#2309) or re-litigate the quarantine lane; the residual expression excludes quarantine. | Separate tracked debt; the lane is #2317's landed design. | Accepted |
| C-006 | pytest.ini remains the only marker registry (PR #2047 + its guard); this mission READS it, never adds a second registry surface. CI_INVISIBLE maps routing-exemption reasons, not marker definitions. | Single canonical authority. | Accepted |
| C-007 | This mission lands GREEN including its own gates: pre-merge dry-run of the residual job selection + the full invariant suite on the final workflow state (gate-unmask-cannot-self-validate discipline). | A gate that first bites post-merge is a regression vector. | Accepted |
Success Criteria (mandatory)
| ID | Criterion | Measurement |
|---|---|---|
| SC-001 | -m unit and -m contract are positively selected by a blocking CI gate | FR-002 job live; FR-001 invariant green; #2034's core claim no longer reproduces |
| SC-002 | Marker→job completeness fails closed | fault-injection evidence recorded (synthetic marker red; de-route red) |
| SC-003 | Orphan baseline at zero (± one recorded intentional exception) | _gate_coverage --check output in the closeout report |
| SC-004 | The WS5 coherence relations hold on live workflows and are guarded | FR-003/FR-005 invariants green; the fixed incoherences enumerated in the PR body |
| SC-005 | No CI-minute regression beyond the residual job's budget | NFR-002/NFR-003 measurements recorded |
Assumptions
- The residual set stays green between spec and implement (debbie verified all 17 live orphans + samples pass); any drift is handled by the fix-all-before-gate bar (Decision 5).
quality-gateremains the single blocking aggregator; branch protection semantics unchanged.- The
workflow-scope push path works from this checkout (verify early per C-004).
Out of Scope
- #2283 factors (b) venv-skew doc and (c) producer-conformance sweep → CT7 (#2077).
- Quarantine triage (#2295/#2309); the quarantine lane's design.
- mypy CI-scope expansion + advisory-vs-enforced decision → #1868 follow-up.
- Marker-registry single-sourcing (done, PR #2047).
slow-testspush-only semantics (documented behavior; changing it is a CI-budget decision this mission doesn't own — recorded, not fixed).- Any test-content remediation beyond what the new gates unmask (CT1–CT5 own content classes).