tooling: the workbench boot check refuses a dead port; verify.sh checks tests/

[itsdestin]

Companion to youcoded#384. Nothing here reaches users.

1. The workbench boot check was a false green

node scripts/workbench-boot-check.mjs 5999 — a port nothing was listening on — printed ok for all sixteen routes and exited 0:

ok    parent frame (toolbar)
ok    app · default
...
All 16 workbench routes mount cleanly.

Chrome renders its own ERR_CONNECTION_REFUSED page, which has no "failed to start" text, no #boot spinner, and throws no exception — so all three of the script's probes read clean. CLAUDE.md tells every session to run this after any mock-shim change, so a dead dev server read as a passing app.

Three assertions now stand between a route and ok, each proven to fire on its own:

Assertion	Proof
Preflight HTTP request to the port, before Chrome is discovered or launched	dead port → exit 2, nothing is serving the workbench on port N, naming run-workbench.sh
Page.navigate's errorText + the main document's HTTP status	a server answering 500 → server answered HTTP 500 for the page
#root, which index.html ships inline	a live server that is not the workbench → #root is not in the DOM — this page is not the workbench

No false red: against a real workbench, 16/16 routes, exit 0.

Guarded by scripts/workbench-boot-check.test.mjs — two cases, no Chrome needed (the preflight now runs before the browser launches) — and a new Workspace CI step. CLAUDE.md's route count corrected 12 → 16.

2. verify.sh

• Runs tsc -p tsconfig.tests.json beside the src one and prints how many test files are still excluded (types in tests/ (tsc --noEmit, 57 file(s) still excluded)), so the debt is visible on every run. Its SCOPE header no longer claims the test tree is unchecked.
• Warns loudly when desktop/node_modules is a symlink. youcoded#384 fixes the test-runner half (Vite denying the resolved path), but npm ci and Gradle's bundleWebUi still follow the link and empty the shared copy, and a green run must not imply the setup is fine.

3. Roadmap and reports, matched to what was measured

• Closed: the boot-check entry, "the desktop test tree is neither type-checked nor linted", and the Denied ID …?inline entry. Two investigations archived.
• Re-scoped rather than closed: the entry naming subagent-view, mcp-startup-wiring and project-watcher as the suites that flake under parallel load. None of the three failed in 27 full local runs (1 alone, 6 concurrent, 4 pinned to 4 cores, 2 × 8 concurrent). The four suites that did fail at 8-way concurrency were different files, and they are fixed in #384. A trigger nobody has reproduced is not a fixed bug, so the entry stays open with what was actually measured.
• New entry for the 57 test files still excluded from the typecheck.
• Fixed-sleeps investigation re-measured: 108 → 102, and the "mcp-startup-wiring blows even the 30 s budget at eight concurrent suites" claim did not reproduce in any of the 27 runs.

Items 6 and 7 of the overnight list needed no change and are recorded in the handoff instead: the audit-staleness hook fix landed on master as 5598f69 (the shared checkout was 98 commits behind, which is why it looked dead — proven firing by running the hook), and node scripts/audit-anchors.mjs is green at 394/394 with all four "already done" roadmap entries already closed.

Verification

node scripts/roadmap-check.mjs --quiet      # exit 0, structure clean, 100 claims / 0 broken
node scripts/audit-anchors.mjs --no-diff    # anchors 394/394, MAP paths 345/345, MECHANICAL PASS: OK
node --test scripts/workbench-boot-check.test.mjs .claude/hooks/context-inject.test.mjs   # 14 pass, 0 fail
bash scripts/verify.sh <worktree> --full    # all six checks PASS

🤖 Generated with Claude Code

https://claude.ai/code/session_01T9SRtMoZJNF4sJrodQa1N1