Reanimation backlog: REV-001, REV-002, the cline-dev-tara merge, the dev-loop fix, and REV-008

[richardosborne14]

Scope note. This PR spans 29 commits, not one task. origin/cline-dev had not been pushed to since d1dacbd, so opening a PR from the REV-008 branch diffed against a base four months stale and swept up everything in between. The body below was originally written for REV-008 alone; it has been corrected to describe what actually lands.

Everything on local cline-dev since d1dacbd. Merging this brings the remote up to date with work that was verified locally but never pushed.

What's included, oldest first

Work	Commits	Status
Tara's January branch — embedded template system (TASK-009), StyleTokens (STYLE-001), ElementConfigs (STYLE-002)	16	merged via 45ac274, see MERGE-NOTES-cline-dev-tara.md
REV-001 — the reanimation commit, restoring editor build + typecheck	1	✅
Revival plan — viability review and phase 12–20 task specs	1	✅
REV-002 — restore the Electron test harness	1	✅
Dev-loop fix + headless debugging tooling (1502581)	1	✅
REV-008 — dev loop hardening and verification debt	7	✅
Workflow rule: work directly on cline-dev	1	✅

Suite at the tip: 705 specs, 0 failures. typecheck:editor clean.

---

REV-008 — trustworthy dev loop, visibility, and merge verification debt

The most substantial piece, implementing REV-008. Suite 540 → 705 specs.

Stream A — committed build artefacts

186 generated files (~24 MB) untracked: the three entry bundles, their split chunks, the Monaco workers, and the file-loader SVGs webpack emits into the package root. These were committed build output living in src/, indistinguishable from source — the root enabler for the black-window bug fixed in 1502581, and the reason every dev run left the working tree dirty. The vendored parse-dashboard bundles stay tracked; nothing in this repo rebuilds those.

npm run check:artefacts plus a CI workflow stop it coming back.

The packaged app had never launched

Verifying Stream A from a genuinely clean clone — which the task insisted on — surfaced two pre-existing defects with nothing to do with tracked bundles:

• dispatcher.getOwner is not a function. react/react-dom are webpack externals, so the packaged app require()s them at runtime. react/index.js picks dev vs production from process.env.NODE_ENV, unset in a packaged Electron app, so development react was paired with the production react-dom the build had just produced. React 19's shared internals differ, and it throws before first paint. This is the same error the REV-008 background attributes to a stale bundle in dev — but here in a fresh production build, which means the packaged app was equally broken all along.
• Cannot find module 'split2'. split2, crelt, style-mod and w3c-keyname are externalised (hoisted to the root node_modules as transitive deps of the file: workspace packages @noodl/git and @noodl/noodl-core-ui) but never declared by noodl-editor, so electron-builder did not collect them.

The NODE_ENV fix nearly shipped as a no-op: webpack's DefinePlugin constant-folds the dotted process.env.NODE_ENV, so the branch was dropped as dead code. cdp health against the packaged build caught it — which is the point of having it. Computed member access survives.

A clean clone now produces a .dmg whose app renders, verified via CDP.

Stream B — visibility

• cdp.js takes --target=editor|viewer, so the preview window is inspectable. Requesting a viewer that isn't running is an error, never a silent fall back to the editor.
• cdp services probes every dev port plus the freshness of the viewer/cloud-runtime watch builds. Both IP stacks are probed — the dev server binds localhost, which resolves to ::1 first on macOS, and an IPv4-only check reports a healthy server as down.
• cdp click / cdp type drive Input.dispatchMouseEvent and Input.insertText rather than synthetic DOM events, so React handlers and focus behave as they do for a user.
• dev-debug.js attaches over CDP as each page appears and writes [renderer:exception] into .logs/dev.log, so a boot crash is recorded without anyone attaching in time to see it.
• The main process is debuggable at last: --inspect has to reach the Electron binary, so start:_dev goes through a launcher. npm run dev:debug -- --inspect-main.
• main.js gains the ELECTRON_RUN_AS_NODE guard test.js already had. Without it the app exits 0 with no output when launched from a VS Code terminal — indistinguishable from a successful launch.

Stream C — screenshots

Both DEBUG-INFRASTRUCTURE.md and the run-editor skill now say CDP-first, and that a black screencapture frame is a macOS permission belonging to the capturing app, not a crashed editor.

Stream D — merge verification debt

• D1. The embedded hello-world template works, and is now covered — including the combination the merge created by hand and that had never executed (tara's template flow plus runtimeVersion = 'react19'). Style-token injection was also meant to be verified here; it cannot be, because nothing ever constructs StyleTokensInjector. See below.
• D2/D3. GroupConfig keyed on net.noodl.visual.group, which is not a node type. Deleted rather than repointed, and tara's ImageConfig not ported: activating either would stamp var(--surface), var(--space-4) and friends onto new nodes, and nothing defines those.
• D4. All six stranded specs converted from Jest to Jasmine and enabled one at a time. Beyond dropping the @jest/globals import: toHaveLength → toHaveSize, it.each → a generated loop, and jest.mock → spyOnProperty (assigning ProjectModel.instance is not viable — its setter registers the value with NodeLibrary).

Two specs were not merely stranded but wrong, which is what six months of never running buys you:

• UBAConditions built its fixture from flat dotted keys ('auth.type': 'bearer'), but evaluateCondition resolves fields with getNestedValue, which needs a nested object. Six specs failed on first run; the implementation is correct, the fixture was not.
• ProjectCreationWizard defined its own copies of getStepSequence and isStepValid and asserted against those, with a comment saying "if the context logic changes, update both files". It never imported WizardContext, so it could not have caught any regression. The two helpers are now exported and the copies are gone.

Contrary to the merge notes, ElementConfigRegistry.test targets the registry that survived the merge, not tara's class API, and needed no rewrite.

Follow-up opened

REV-009 — the style-token system was never connected to anything. StyleTokensInjector is never constructed, so no tokens reach a preview; the ten tokens that did ship use a different vocabulary from the one ElementConfigs references (which exists only in the phase-9 design docs); and ElementConfigs is live, persisting fontFamily: var(--font-sans) and four more unresolvable values into every new Text node. Needs a product decision, not a mechanical fix.

Verification

• npm run test:ci — 705 specs, 0 failures, exit 0
• npm run typecheck:editor — clean
• Clean clone → npm install → npm run build:editor → .dmg + .app; launched with cdp health → reactMounted: true
• Clean tree → npm run dev:debug -- --inspect-main → cdp health reactMounted: true, cdp services all up, and the working tree stays clean after a dev run

---

Workflow change

.clinerules now says to work directly on cline-dev — single developer, so a task branch plus a PR adds a merge step and a review round-trip with nobody on the other end. The not-yet-started phase-12 specs have their Branch: field and "open PR" checklist items updated to match. This is expected to be the last PR of this kind.

🤖 Generated with Claude Code