Jaw Dev Testing

Testing & QA

Balance: ~40% Backend/API, ~40% Frontend/E2E (Playwright), ~20% Cross-cutting (CI, Security, TDD, Coverage) -- directional guidance, not a hard ratio. Scope: test harnesses, fixtures, mock policy, runners, Playwright, CI gates, coverage. Root-cause analysis and debugging playbooks → dev-debugging.

• Review/test boundary and test adequacy findings: see dev-code-reviewer.
• CI pipeline ownership and deployment verification: see dev-devops.
• Data pipeline testing and ETL validation: see dev-data.
• Design direction context for rendered verification: see dev-uiux-design. This skill activates by change surface when work needs verification depth, regression coverage, or a reproducible test harness.

C0/C1 work (small local patches): See dev §0.0 Work Classifier + §0.1 Patch Fast-Path before reading references.

dev is canonical: dev §0.2 Rule Classes, §3 Verification Gate, and §5 Safety Rules apply to all work governed by this skill.

Modular References

When tests depend on current external API behavior, provider docs, CI service behavior, test-environment versions, dependency audit evidence, or recorded mock/fixture sources, read the active search skill and follow its source-fetch and evidence-status rules.

1. Test Strategy

1.1 Models

1.2 Recommended Trophy Distribution

1.3 Risk-First Priorities

1. auth / session / permission boundaries
2. money movement, quota, credits
3. data mutation and irreversible actions
4. file upload / parsing / external webhooks
5. shared API contracts used by frontend clients
6. error paths, retries, rollback behavior

1.4 Harness Selector

1.5 General Rules

• Write tests for new features, bug fixes, refactors, and behavior changes.
• Prefer one behavioral concern per test.
• Use factories / builders for setup; avoid repeated inline blobs.
• A fast real dependency beats a mock. A mock beats an untested branch.
• If the failure is mysterious, delegate methodology to dev-debugging, then return here for the regression harness.
• STRICT (TEST-ANTI-FLAKE-01): A time-based flake is a bug. Do not use sleep-based synchronization, retry-as-fix, or green-on-retry acceptance without a deterministic cause and harness correction.
• Verification depth follows dev §3 DEV-VERIFY-FLOOR-01; CRUD per-operation negative coverage is owned by references/core/crud-test-matrix.md.

1.5a Limited-Oracle / Score-Objective Evaluation

When the real evaluator is scarce, paid, rate-limited, or opaque and local tests are proxy metrics for a score/objective, apply §9.5 (single owner of GATE-ORACLE-VALIDITY-01, GATE-PREFIX-HORIZON-01, GATE-INVARIANT-EV-01, GATE-HOLDOUT-LEAKAGE-01, GATE-AGREEMENT-STATS-01). Pairs with dev-pabcd §10 Optimization-Loop Meta-Rules.

1.6 Property-Based & Mutation Testing (verified 2026-07-02)

• Vitest 4 is the current runner baseline: Browser Mode is stable (visual regression via toMatchScreenshot, Playwright trace generation, expect.schemaMatching).

2. Backend & API Testing

Deep reference: references/backend-testing.md

2.1 Coverage Map

2.2 Mock Strategy Hierarchy

real deterministic dependency
→ Testcontainers / ephemeral infra
→ recorded responses / thin fake
→ manual stub / fake
→ framework mock as last resort

2.3 Service & API Patterns

Mock dependencies at service boundaries. Use Supertest/httpx for route-level integration tests. Match response envelope shape from backend contracts.

2.4 Database Truth with Testcontainers

Use a real database when verifying migrations, transactions, unique constraints, foreign keys, query translation, and performance-sensitive SQL. Use Testcontainers for real DB truth in correctness-sensitive persistence tests. Start container in beforeAll/fixture setup, capture connection URI.

2.5 Fixture / Seed Synchronization

• Prefer builders / factories over copied JSON snapshots.
• Keep shared contract examples in fixtures/contracts/ or equivalent.
• Seed data should expose stable IDs used by Playwright smoke flows.
• If frontend mocks drift from backend fixtures, write or update a contract test first.

3. Contract Testing

Contract tests protect the frontend↔backend boundary. They sit between API tests and browser tests. Rule: Playwright proves the experience. Contract tests prove the shared shape.

3.1 Contract-Stable Surface

• response envelope: success, data, error, meta
• error taxonomy: HTTP status + machine-readable error.code
• pagination fields, auth headers, cookie behavior
• requestId propagation
• nullability, timestamps, enums, money serialization

3.2 Contract Options

3.3 Consumer Contract — TypeScript (Pact / PactV4)

PactV4 (aliased Pact) is the current interface and supports Pact Specification v4; treat PactV3 as the legacy spec-v3 API. Workflow:

1. Define interaction: provider state + request + expected response (matchers for flexible matching)
2. Execute test against Pact mock server
3. Assert consumer expectations
4. Pact file auto-writes to pacts/ → publish to broker → provider verifies

See references/backend-testing.md for a full example.

3.4 Schema Verification

Use schema-based API testing (Schemathesis) to verify OpenAPI/GraphQL contract compliance. (Dredd is legacy/inactive — do not adopt for new projects.)

3.5 Rules

• Contract tests are strongly recommended for parallel FE/BE, public APIs, and cross-team contracts.
• E2E success does not replace provider verification.
• Store golden examples near the contract, not inside one app only.
• If the shape is intentionally breaking, update the contract first, then all consumers.

4. Playwright Browser Testing

Use Playwright after API and contract tests are already trustworthy. Browser tests should validate rendered flows, accessibility-critical interactions, and real integration seams that lower layers cannot prove alone. Helper Scripts Available:

• scripts/with_server.py - Manages server lifecycle (supports multiple servers) Run scripts with --help first — treat as black boxes to avoid context window pollution.

4.1 Decision Tree: Choosing Your Approach

User task → Static HTML? → Read file → find selectors → write Playwright script
         → Dynamic app? → Server running? → No: `python scripts/with_server.py --help`
                                           → Yes: Recon-then-action (navigate → screenshot → selectors → act)

4.2 Example: Using with_server.py

# Single server:
python scripts/with_server.py --server "npm run dev" --port 5173 -- python your_automation.py

# Multiple servers:
python scripts/with_server.py \
  --server "cd backend && python server.py" --port 3000 \
  --server "cd frontend && npm run dev" --port 5173 \
  -- python your_automation.py

4.3 Reconnaissance-Then-Action Pattern

1. Wait for an explicit app-ready signal or locator assertion → 2. Screenshot/inspect DOM → 3. Identify selectors → 4. Execute actions

4.4 Best Practices

• Use bundled scripts as black boxes — run --help first, invoke directly.
• Use sync_playwright() for synchronous scripts; always close the browser.
• Prefer locator-based interactions and web-first assertions: expect(page.get_by_role("button", name="Save")).to_be_visible(), then click() on that locator.
• Prefer user-facing locators, especially get_by_role() with an accessible name. Use get_by_label(), get_by_placeholder(), or get_by_test_id() when role/name cannot express the target.
• Avoid networkidle, hard sleeps, and wait_for_timeout() in tests. Wait on observable app-ready signals, locator actions, or expect() assertions.
• AI-authored tests (DEFAULT): Playwright MCP / Test Agents (planner/generator/healer) are generation-and-repair aids only — final acceptance still requires deterministic locators, web-first assertions, traces, and a human-readable failure artifact.

4.5 Reference Files

• examples/ - Examples showing common patterns:
  • element_discovery.py - Discovering buttons, links, and inputs on a page
  • static_html_automation.py - Using file:// URLs for local HTML
  • console_logging.py - Capturing console logs during automation

4.6 Browser Testing Rules

• Run contract tests and API tests first for broken-data bugs.
• Use Playwright for rendered truth, not as a replacement for service tests.
• Prefer one smoke flow per critical path over many brittle micro-flows.
• If a failure looks like data-shape drift, go back to §2 Backend & API Testing or §3 Contract Testing.

4.7 Native Computer-Use / Browse-Use QA (exploratory tier) (TEST-CU-QA-01)

Browser QA loads dev-frontend for rendered implementation context. Playwright owns deterministic suites; native tools own immediate exploratory proof. QA-TOOL-LADDER-01: start at 1 and state why when skipping:

1. Manager embedded browser (POST …/<targetId>/snapshot for the accessibility tree, …/screenshot, …/act for click/type/scroll/key) for built or locally served web UI.
2. A full real-profile browser for login, extension, or WAF state the embedded browser cannot carry.
3. Desktop or screen-level control for GUI-only flows; keep credential entry human-supervised.
4. agbrowse only for public-URL response-shape proof, never built-UI driving. Use inspect -> act -> re-inspect; when DOM inspection fails, take a screenshot AND read the image back -- a screenshot produced but never read is not an observation. Evidence names the flow, states, result, and screenshots; promote durable flows to Playwright.

5. CI Pipeline Integration

Full workflow templates: references/ci-pipeline.md

5.1 Pipeline Order

quality (lint / typecheck)
→ unit + integration tests
→ contract tests
→ Playwright E2E
→ security scan
→ coverage aggregation + artifacts

5.2 Pipeline Template

Structure CI jobs in dependency chain: quality → backend-tests → contract-tests → e2e

Key configuration:

• concurrency.cancel-in-progress: true — avoid wasted runs
• strategy.fail-fast: false — for matrix builds
• Shard large suites: --shard=${{ matrix.shard }}/N
• Install Playwright deps: npx playwright install --with-deps chromium

See references/ci-pipeline.md for full GitHub Actions and GitLab CI templates.

5.3 Matrix & Parallelization

npx vitest run --shard=1/4
npx playwright test --shard=1/4 --workers=4
pytest -n auto --dist=loadgroup

5.4 Flaky Test Remediation

5.5 Rules

• Do not let Playwright be the only blocking job.
• Contract tests should run before browser tests.
• Upload artifacts for failures: coverage, junit, traces, screenshots.
• Fail the build on broken thresholds, not only test exit codes.

6. TDD Enforcement Mode

When ENFORCE_TDD=true is set in project instructions or explicitly requested, this section becomes mandatory.

6.1 RED → GREEN → REFACTOR

1. RED — write the failing test first and verify it fails for the right reason.
2. GREEN — write the minimum implementation to pass.
3. REFACTOR — clean up after green, then rerun the affected suite.

6.2 Self-Audit Checklist

6.3 Vertical Tracer-Bullet TDD

Prefer one behavior test → minimal implementation → next behavior. Slice by something a user, caller, or consuming module can observe, not by horizontal layers such as "DB", "API", then "UI". Assert through public interfaces and durable contracts. Retire shallow scaffolding tests when a stronger interface or acceptance test covers the same promise.

6.4 Default Style

6.5 Boundary with dev-debugging

• dev-testing owns the regression harness and enforcement loop.
• dev-debugging owns root-cause methodology once a failure is mysterious or multi-layered.
• After dev-debugging isolates the cause, come back here to lock it in with tests.

6.6 AI-Assisted Development Regressions

When an AI writes and reviews its own code, it carries the same assumptions into both steps. Automated tests break this feedback loop.

Common AI Regression Patterns

Regression Naming Convention

Name regression tests with BUG-R{N} convention. Assert all required fields with a loop.

Sandbox-Mode API Testing

When the project supports a sandbox/mock mode, use it for fast DB-free regression testing:

• Force sandbox mode in test setup: process.env.SANDBOX_MODE = 'true'
• Assert sandbox responses match the same contract as production responses.
• Treat sandbox/production parity as a high-priority regression target.
• In sandbox/spike mode, write tests for bugs found — coverage grows organically. For production refactors, see §1.5 (tests required for behavior changes).

6.7 Test-Induced Production Defense Detection

Rule: Do not add production defensive code solely to satisfy unrealistic tests. A production guard is allowed only when the invalid state can occur at a real boundary or represents an explicit domain rule.

Required questions before adding a guard:

1. Is the input from an untrusted boundary? → If yes, validate at that boundary
2. Can this state happen in production? → If no, fix the test
3. What contract allows this value? → Cite schema/type/domain rule
4. Would this hide a real bug? → If yes, fail fast instead

Banned patterns: process.env.NODE_ENV === "test" branches, silent fallbacks for impossible internal state, making required types optional for mocks, exporting internals only for tests.

Allowed guards: Boundary validation (process/network/user/file boundary), backward compatibility (documented old schema), security checks, domain invariants, observed production bug regressions, external dependency adapters.

7. Accessibility Testing

Component Level

• jest-axe / vitest-axe: run axe-core on rendered components

  import { axe, toHaveNoViolations } from 'jest-axe'
  expect.extend(toHaveNoViolations)
  expect(await axe(container)).toHaveNoViolations()
  

Page Level

• Playwright a11y assertions:

  import AxeBuilder from '@axe-core/playwright'
  const results = await new AxeBuilder({ page }).analyze()
  expect(results.violations).toEqual([])
  

CI Pipeline

• Gate order (verified 2026-07-02): component axe (jest-axe/vitest-axe) → page axe (@axe-core/playwright) → keyboard/focus/manual checks. Blocking gate = zero serious/critical axe violations + targeted manual checks.
• Lighthouse a11y score (≥90) is an advisory route-level smoke signal, NOT the blocking gate — Lighthouse itself separates automated scoring from required manual checks.
• Pa11y: page-level scanning for WCAG AA violations
• Run a11y tests on EVERY page route, not just the homepage

Observability Verification

Verify trace propagation in integration tests. Assert that spans appear for critical paths. Check structured log format matches the schema in dev-backend/references/core/observability.md.

8. Security Testing

→ Delegated: threat modeling and secure design policy belong to dev-security. This section covers the automated test hooks and CI gates that enforce those rules.

8.1 Minimum Security Stack

fast local checks
→ Semgrep / CodeQL gate
→ dependency audit
→ auth / validation regression tests

8.2 Dependency Scanning Commands

npm audit --audit-level=high
pip-audit --strict --desc

8.3 Semgrep Gate

The returntocorp/semgrep-action wrapper is deprecated (stated by the repo itself) — run Semgrep natively in CI:

semgrep:
  runs-on: ubuntu-latest
  container: semgrep/semgrep
  steps:
    - uses: actions/checkout@v4
    - run: semgrep ci --config p/default --config p/javascript --config p/typescript --config p/python

(Open-source alternative engine: Opengrep, the LGPL-2.1 community fork — see dev-security/references/static-analysis.md.)

8.4 Security Regressions

Test missing auth (expect 401) and verify error.code matches contract for every auth-protected endpoint.

8.5 Rules

• dependency audit in CI
• Semgrep or equivalent SAST
• auth / permission regression tests
• validation tests for malicious or malformed input
• a blocking rule for high / critical dependency findings

9. Coverage & Quality Gates

9.1 Suggested Thresholds

These are project/risk-based, not universal minimums. Adjust for your context.

9.2 Outcome Metrics

9.3 Coverage Workflow

1. generate coverage reports

   npm test -- --coverage
   npx vitest run --coverage
   pytest --cov --cov-report=xml
   

2. review by priority: auth, payment, mutations, upload, contracts first
3. write targeted tests for the gaps
4. publish artifacts and fail the merge when thresholds drop

9.4 Quality Gate Checklist

• focused unit / service tests
• API integration tests for changed routes
• contract tests for shared payload changes
• Playwright smoke for critical rendered journeys
• security scan / dependency scan
• coverage thresholds and diff coverage
• CI artifacts uploaded for failure analysis

9.5 Limited-Oracle / Score-Objective Evaluation

Use these rules when the true evaluator is scarce, paid, rate-limited, or opaque and local checks are only proxy metrics for a score/objective. PABCD loop response to repeated candidate deaths is owned by dev-pabcd §10 Optimization-Loop Meta-Rules.

• STRICT (GATE-ORACLE-VALIDITY-01): When the true evaluator/oracle is rate-limited (limited submissions, paid runs) and local metrics are proxies, evaluator validity is a prerequisite gate. Before trusting the proxy for accept/reject, quantify historical divergence: cases where the proxy said better/equal but the oracle said worse. A proxy with known optimistic bias must not be the sole acceptance evidence.
• DEFAULT (GATE-PREFIX-HORIZON-01): Replay-based evidence (recorded logs, scripted opponents) is prefix-valid only; it stops being valid when the candidate diverges from the recorded trajectory. Candidates that diverge early need live adversarial evaluation through a modeled opponent/environment, not replays. State the divergence turn/point when citing replay evidence.
• DEFAULT (GATE-INVARIANT-EV-01): Every hard invariant in an acceptance gate (a metric that must not regress) needs an expected-value justification: protected value versus the candidate-space it vetoes. If a hard invariant vetoes 3+ consecutive candidates that target strictly larger gains, downgrade it to a soft cost and re-justify or remove.
• DEFAULT (GATE-HOLDOUT-LEAKAGE-01): Fixed evaluation sets become training data under adaptive reuse. Rotate or quarantine repeatedly used sets, keep a blind final set, and treat repeated score-driven tuning on the same instances as leakage risk. Grounding: Blum and Hardt, "The Ladder: A Reliable Leaderboard for Machine Learning Competitions" (arXiv:1502.04585), and Dwork et al., "The reusable holdout: Preserving validity in adaptive data analysis" (arXiv:1506.02629).
• HEURISTIC (GATE-AGREEMENT-STATS-01): For proxy-vs-oracle validation, report agreement statistics that expose error shape: sign-discordance, bias, and worst-case error. Correlation alone is not enough; Bland-Altman style agreement analysis is the grounding model.

Grounding: observed in a 14-discard optimization plateau where a prefix-only replay gate and a hard draw-protection invariant locked a 3.5/8 score.

10. Pre-Flight Test Checklist

10.1 Change-Type Routing

• pure business logic change → add / update unit or service tests
• API or middleware change → add / update API integration tests
• shared frontend↔backend payload change → add / update contract tests
• rendered user flow change → add / update Playwright smoke coverage
• auth / upload / billing / external integration change → add security or edge-case regression coverage

10.2 Harness Readiness

• fixtures are deterministic and reusable
• real dependencies are used where correctness matters
• Testcontainers are used for DB truth, not mocked SQL
• external APIs are mocked or recorded intentionally, not accidentally called live
• ENFORCE_TDD requirements were followed if enabled

10.3 Contract & Data Integrity

• response envelope remains stable or contract was updated first
• error codes are asserted, not only HTTP status
• requestId, pagination, and nullability are verified where relevant
• frontend fixtures do not drift from backend examples

10.4 CI & Reporting

• relevant CI jobs exist and are actually executed
• sharding / matrix choices match project size
• flaky failures were investigated instead of blindly retried
• coverage / junit / trace artifacts are available on failure

10.5 Final Rule (risk-tier)

Verification intensity follows the work class (dev §0.0 / references/core/crud-test-matrix.md): for C2 UI work, one focused smoke (manual click-through or one Playwright run) plus targeted checks IS a complete story; for C3/C4 or release-sensitive work, a single smoke is not enough — run the affected suites and required negatives. Manual/Playwright smoke is a risk-tier rule, not a universal blocker.

unit / service
→ API integration
→ contract verification
→ Playwright smoke
→ CI gate + coverage + security scan

Patch Integrity Gate (TEST-PATCH-INTEGRITY-01, DEFAULT)

Source: sol research (SWE-bench containerized evaluation, addyosmani/agent-skills).

An agent that obtains green by weakening tests has not fixed the bug. Before claiming implementation complete:

1. Baseline: record which tests fail and their failure signatures BEFORE any production code change.
2. Implement: write the fix/feature.
3. Re-run originals: execute the ORIGINAL test suite (not the modified version) against the patched source. All baseline failures must now pass.
4. Classify test changes: every test/config change is required (new test for new behavior), suspicious (deleted assertion, lowered threshold, added skip, reduced coverage exclusion), or unrelated.
5. Justify suspicious changes: each suspicious change needs a stated reason. "The test was wrong" is valid only with evidence of the original test's incorrectness.

Red flags that trigger escalation:

• Deleted assertions without replacement
• Snapshot updates without visual/behavioral verification
• Coverage exclusions added in the same PR as the fix
• @skip or .skip() added to failing tests
• Threshold reductions (e.g., coverage 80% → 60%)
• Type assertion suppressions (as any, @ts-ignore) in test files

TDD Evidence Contract (TEST-TDD-EVIDENCE-01, DEFAULT)

When TDD is claimed, durable evidence must show:

• RED: failing test name + failure message (before production code)
• GREEN: same test passing (after production code)
• REFACTOR: full affected suite passing (after cleanup)

A TDD claim without RED evidence is not TDD.

Acceptance-Row Reachability (TEST-ROW-REACHABLE-01, DEFAULT)

Every row of an acceptance-criteria table must have a constructible precondition. Before writing a row, ask: is there a call path that reaches this state? Does an earlier guard consume this condition first, so the branch under test is never entered? Does an operation that produces this value actually exist?

An unreachable row is decoration, not verification — the implementer tries to write that test, cannot, and quietly drops it. This applies C-ACTIVATION-GROUNDING-01's requirement (every conditional path names its activation scenario) to each row of the acceptance table, not just to the plan's prose.

Common unreachable shapes:

• The row asserts a rejection that an earlier, broader rule already rejects — the specific guard is never exercised, so the test passes even if that guard is absent.
• The row needs a state the public API cannot produce (no operation creates it).
• The row asserts on a call-site argument the function never receives.
• Two rows in the same table are mutually exclusive on the same tree (for example "the gate reports a failure here" together with "the gate exits 0 overall").

Test Oracle Integrity (DEFAULT)

Three narrow contracts that stop false-green. No gate enforces any of them; the reviewer is the only check.

TEST-PROMPT-SEAM-01. Do not assert on prose. A test may read a document only when it extracts a value and compares it against a value from another source. Asserting that a phrase exists in a file is a violation no matter how many files you read or what the test's header says it is for.

• Forbidden: assert.match(readFileSync(".../SKILL.md"), /prefer rg first/i) — one source, phrase existence, breaks on harmless rewording, proves no behavior.
• Allowed: parse the frontmatter description, pull a token out of it, and assert that token appears in the runtime constant it is supposed to mirror — two sources, values compared, breaks only when they genuinely disagree.
• Outside this rule's scope, and allowed: asserting on non-prose values — version pins, license names, runtime output, CLI stdout, file existence.

The rule was written after two such tests were deleted rather than repaired, and the reason repair failed is the reusable part. Three designs were tried: a structured metadata field in each document compared between them (nothing at runtime reads such a field, so it becomes a third source of truth that passes whenever both copies are wrong together); promoting the prose to a behavioral test (the behavior was already covered, so this duplicated coverage while appearing to offset the deletion); and promoting a doctrine statement the same way (no runtime branch implements it, so there is nothing to compare against).

What that costs is stated plainly rather than hidden: the affected doctrine has no automated consistency check and is a human-review item. When prose has no counterpart in code, a test that reads it can only check that the words are still there. Deleting it removes a false green; inventing a second document to compare it against removes nothing and adds a lie.

TEST-ORACLE-INDEPENDENCE-01. Never derive the expected value from the code under test.

• Forbidden: assert.equal(fn(x), fn(x)); building the expectation with a helper the code under test also uses; refreshing a snapshot from current output and calling that verification.
• Allowed: hardcode the expectation in the fixture, or compute it by an independent route — a second implementation, a hand-worked example, an external spec.

TEST-PRECEDENCE-FIXTURE-01. When testing override / default / fallback, the three values must all differ — otherwise the test cannot tell which path ran.

• Forbidden: override "x", default "x", fallback "x" — every branch passes.
• Allowed: override "from-flag", default "from-config", fallback "builtin", with each case asserting the specific one.

A regression test should FAIL when the defect is reintroduced. Confirm it once by mutation — break it, watch it go red, restore it, watch it go green. A regression test that has never been seen red is an assertion about the test, not about the defect.

CI Green Discipline (TEST-CI-GREEN-01, STRICT)

The latest HEAD is the source of truth for whether CI is green.

1. Inspect before editing. Read the failing job's log and its artifacts. A failure you have not read is a failure you are guessing about.
2. Make the minimal correct fix for that failure.
3. Re-watch the latest HEAD, not the run you were looking at when you started.

Never blind-retry a failed job, and never push another change without new failure evidence. Both produce the same illusion — activity that looks like progress while the failure is unexamined. Re-running to measure a failure RATE is a different act and is allowed; see references/ci-pipeline.md §5.2.

Exploratory QA Tier (TEST-CU-QA-01, DEFAULT)

Deterministic browser test suites and exploratory QA are different tools for different jobs, and the split is worth stating because collapsing it wastes both:

• Deterministic suites own regression guarding. They are the thing you keep.
• Exploratory QA through whatever browser or screen-control capability the runtime offers owns immediate proof that a surface built right now actually works.

Protocol: inspect → act → re-inspect. When DOM inspection is unavailable, fall back to screenshots and read the image back — a screenshot produced but never read is not an observation.

Evidence names the flow, the states traversed, the result, and the artifacts. Promote any flow that must stay guarded into a deterministic test; exploratory proof does not accumulate, so a flow left only in QA is a flow that will regress unnoticed.

The tool ladder for this tier is QA-TOOL-LADDER-01 in jaw-dev — start at the runtime's embedded browser and escalate only with a stated reason.