Ship Nod in the app bundle and make it launchable where its ramp is on

[scgopi]

Final integration for GraphCode Nod, after #583 and #588–#592: Nod now ships in the app bundle and is launchable on installs where its beta ramp is on.

Change	Detail
Bundling	make build-nod packages NodRuntime; release-dmg copies it to Contents/Helpers/nod and signs graphcode-nod, copilot-runtime and runtime.node with NodRuntime/packaging/entitlements.plist (JIT)
Launchable	.nod.isSpiked = this process could launch it (ramp flag + runtime present). App, daemon and CLI install the check at startup; the test host does not, so hosting tests don't depend on the Mac's own install
Ramp at launch	The app publishes nod/ramp.on from the cached ramp, so a fresh beta install doesn't wait on the fetch
Lineage	--inherit <brief> on fresh launches only; composite children launch --unattended
One state dir	The chat pane uses NodRuntimeLocator.stateDirectory; the app's duplicate helper is removed

Still owed, in parallel loops: mounting the graphcode MCP server in both engines (NodMcpMount), and the chat pane's graph slots, delegate actions and card wiring (NodWiring).

Evidence

RED: NodLaunchTests.theLineageBriefRidesAFreshLaunchOnly on origin/main -> compile failure, nodArguments has no fresh: parameter and never passes the lineage brief
GREEN: xcodebuild test (private DerivedData) -> exit 0, 2192 tests in 251 suites passed
GREEN: make build-nod, then bun scripts/smoke.ts copilot with the packaged graphcode-nod and both SDK platform packages hidden -> exit 0, hunk accepted, ask allowed, steer, goal check holds 2/2
REGRESSION: graphcode-cli and graphcoded build -> exit 0; make check -> 0 errors; swift-portable test -> exit 0

[scgopi]

Independent review — ✅ no blockers (would approve; GitHub won't let the author's account approve its own PR, so this is posted as a comment)

Reviewed at 9539b55f against origin/main 9ff1442c. Nothing here blocks the merge. The notes below are follow-ups.

What I checked

Area	Verdict	Evidence
make build-nod	✅	package.sh <out> bun-darwin-arm64 writes graphcode-nod, copilot-runtime, runtime.node and claude flat into .build/nod, which matches the three names the signing loop expects. Arm64-only is consistent with release-dmg
Bundling into Contents/Helpers/nod	✅	ditto keeps modes and mtimes. DaemonBootstrap.bundledNodDirectory resolves Contents/Helpers/nod from Resources/bin, and installNodRuntime copies the whole directory and clears quarantine
Signing (Developer ID branch)	✅	The three Nod binaries are signed with hardened runtime, a timestamp and the JIT entitlements before the outer app is sealed. In the ad-hoc branch, --deep re-signs them without hardened runtime, so JIT isn't needed there
Bundled Claude binary	✅ (not run)	claude-agent-sdk-darwin-arm64/claude is already Developer ID Application: Anthropic PBC (Q6L2SF6YDW) with the runtime flag and a secure timestamp, so leaving Anthropic's signature in place should pass notarization. No signed or notarized release-dmg run is in the evidence, so this part is still unproven
isSpiked hook	✅	Installed in all three shipped processes, in the right order: app init (after SupportDirectory.prepare, before the lazy static store), graphcoded (straight after prepare) and the CLI (before parse). The check runs lazily on every call, so the daemon picks up a ramp flag or runtime that appears after it starts. No other executable hosts loops. The test host skips it through XCTestConfigurationFilePath
Ramp flag at launch	✅	publishNodFlag(isEnabled(.nod)) reads the cached configuration, or the default (beta 100 / stable 0), synchronously. It also removes the flag when the ramp is off, so the kill switch still applies from the next launch
--inherit fresh-only, --unattended for composite children	✅	resumeArguments passes fresh: false. The no-prompt and prompt paths both default to fresh. The runtime also drops --inherit under --resume (main.ts:112, runtime.ts:109)
NodStateDirectory removed	✅	No references are left. The pane and the launcher now share NodRuntimeLocator.stateDirectory, with the same uppercase uuidString path
bun test + typecheck ❌	Not this PR	src/mcp/daemon.ts(103) TS2322 came in on main with the graph-layer merge (7bdb150f). This PR doesn't touch NodRuntime/, and #601 fixes it (event[key]!)

Do the tests prove what they claim?

Test	Fails on main?	Notes
theLineageBriefRidesAFreshLaunchOnly	✅ Yes	It doesn't compile on main (there's no fresh: parameter), and its logic would fail anyway because main never reads lineage.briefPath. It covers both the fresh-only rule and the composite-child --unattended
aTimedLoopIsUnattended… (edited)	✅	It now gets the brief from lineage, not an argument
nodHostsNothingWhereItCannotLaunch	⚠️ Passes on main too	It pins the uninstalled default, which is correct. No test exercises installAvailability() or its truth table (ramp on/off × runtime present/absent). That's understandable, since it's process-global state in a parallel suite, but the PR's core switch has no test
theStateDirectoryIsTheUppercaseNodeID	Passes on main	It's a refactor guard and doesn't claim RED

Non-blocking follow-ups

1. The pane's own launch skips the new rules. GhosttyTerminalView.nodLaunchPrefix builds argv without lineage: no --inherit, and unattended only for .timeBased. Its doc says it uses "the same nodArguments the daemon launches with", but that's no longer true. Any Nod loop the pane starts itself won't get the brief. That includes turn-based forks, which only the pane ever starts, and a pane launch that beats the daemon's launch. It was already missing on main for --inherit; the composite-child --unattended difference is new. The pane is the NodWiring work, so it's probably best fixed there.
2. Size. .build/nod is about 360 MB (claude alone is 228 MB). Every DMG and update carries it, including stable, where the ramp is 0%. DaemonBootstrap also copies it into each workspace's bin. It's worth deciding this on purpose before stable.
3. Prove the release path once with make release-dmg SIGN_ID=… NOTARIZE=0, then codesign --verify --deep --strict and a notarytool submit, before the next beta is cut. That's the only way to confirm the claim about the Anthropic-signed nested binary.
4. macos and the Windows jobs were still pending when I posted this.