# Unified application verification Implementation history for `codex/define-app`. These are scoped results recorded during development, not a substitute for final release status. The exact-head checks and reviewer discussion are retained in [PR #23](https://github.com/lakam99/redweb/pull/23). ## Starter migration (2026-09-01, Windows / Node 22.21.0) - `verify:starters:lifecycle:coverage`: 72 unit and real HTTP/WebSocket/process tests pass. `Application.js` and `StartupCleanup.js` are 100% covered in statements, branches, functions, and lines. - The CLI templates, `SourceInspector`, and `Documentation` reach all-four 100% with 80 unit/source-repair integration tests. The narrower unit-only command does not cover external template inspection; include the action-reference and doctor-source suites when reproducing this scope. - `verify:starters:source-coverage` passed for all six original TypeScript applications, separately from compiler-generated JavaScript coverage. The retained run is `coverage/starter-source/a011d72d-1450-409f-935a-d21623c582a6/`. Its summary records input/report hashes and the actual source inventories. - `verify:starter:browser:coverage`: 30 tests pass, including headed counter, chatroom, and multi-page site interactions from compiled/source-removed apps. - `verify:dashboard:coverage`: 27 tests pass, including headed rejected/accepted login, private cards, draft preservation, logout/relogin, and deletion. - The six generated applications pass real-network tests. Their shared process test imports inert definitions, runs each application, closes it via signals or native listener closure, and checks occupied-port startup failure. Windows exercises Node signal events; Linux exercises delivered OS signals. The independent reviewer found no runtime blocker in the migration. The stale generated-helper report is now explicitly historical. The copied `run-app` implementation was removed, not retained beside the library lifecycle owner. An earlier interrupted full-suite run had a browser-control failure before its intended negative assertion. A standalone rerun passed all four working controls and seven intentional faults (63.9 seconds), without changing the control or weakening its assertions. That does not establish the cause of the earlier failure or replace a completed full-suite run. Final regression, package, CI, release-catalogue, and exact-PR-head review remain outstanding. No soak or long fixed-window acceptance test was run for this migration. ## Follow-up regression repair The earlier Linux CI matrix exposed `StartupCleanup` replacing native errors from another JavaScript context with a generic error. The existing owned-listener integration reproduced both failures locally. Using Node's `isNativeError` preserves the original error identity/message without treating arbitrary thrown values as native errors. A real VM-context regression was added; lifecycle coverage is now 73 passing tests at all-four 100%, and all eight owned-listener integration tests pass. The VM/owned-listener suites also pass on Node 18. CI also identified an obsolete test-count assertion in the starter coordinator's input-mutation check. Its expected successful fixture now includes one realtime test and four application-entrypoint tests; failure-on-input-mutation assertions are unchanged. The reviewer caught a Node 18 distinction: `DOMException` is an `Error` there but is not recognized by `isNativeError`. Keeping both checks preserves native abort-reason identity as well as foreign-context errors. A native-process abort regression passes on Node 18 and 22; the lifecycle scope now has 74 passing tests and all-four 100% coverage. The corrected starter coordinator gate passes all 56 tests with all-four 100% coverage of its three coordinator modules. ## Release catalogue checks The prepared package and lockfile are `0.14.0`. The 68-page generated catalogue matches `docs/releases/0.14.0.json`; older release snapshots were not edited. The package dry run includes the unified runtime and shared entrypoint tests, and the release-documentation guard passes. No package was published. The optional docs MCP adapter previously hard-coded an unreleased-channel expectation. Its tests now compare against the loaded canonical channel and exercise both valid channel variants. All seven tests pass, including actual MCP subprocesses and isolated production-only package installation; all three adapter modules have 100% line, branch, and function coverage. ## Broader regression findings The first completed Windows run finished with 174 passing suites and four failing suites (1,984 passing tests, seven failures, five skips). It began before the final source repairs and version update, so it is not final-head evidence. - The standalone-example unit launcher still stubbed `start`. It now explicitly tests `defineApp().run()` and its rejection handler. Both original-TypeScript coverage runs (standard and legacy decorators) pass with unchanged real page behavior checks alongside the separately labelled launcher units. - Current guide version labels were left at 0.13.5. They now match 0.14.0; the candidate catalogue was corrected before publication. No published release snapshot was changed. The documentation unit/coverage and release guard pass. - The HTML load workload timed out awaiting disconnected-session expiry. Its unchanged standalone replay passed: 200 expired renders, 110 clients, and 8,288,824 bytes heap growth. A standalone pass does not establish the earlier timeout's cause. - Server recovery reproduced a server-connection cleanup timeout in isolation. Evidence remains in `coverage/server-recovery-candidate-kFjY7N/`; this failed run is not acceptance. Its limits and workload have not been weakened. Coverage review also added real acceptance for non-live pages with custom socket routes and direct invalid socket-registration validation. Later exact-head test results and any further repairs belong to the PR's final verification record.