Stop duplicate CI runs and surface test runner errors

[egiurleo]

What's broken

PR #807 had a failing build check that had nothing to do with its dependency bumps. The workflow triggers on on: [push, pull_request], so every in-repo branch (including every Dependabot branch) gets two identical runs. For commit 43a9411, the two runs started three seconds apart:

Run	Event	Result
34853206151	push	resolved + downloaded VS Code 1.137.0, passed
34853212259	pull_request	died 290ms into Resolving version...

Same commit, same workflow, opposite results. The duplicate jobs race each other for the same VS Code download and can get throttled. Re-running the failed job passed, confirming the flake.

The failure was also undiagnosable. src/test/runTest.ts caught the error and logged only:

- Resolving version...
Failed to run tests

because console.error("Failed to run tests") discarded err.

Changes

• .github/workflows/ci.yml — limit push to main, so branch pushes are covered once by the pull_request event. Adds a concurrency group so superseded PR runs get cancelled. cancel-in-progress is conditional on pull_request, so post-merge runs on main are never cancelled and every merged commit keeps its own signal.
• src/test/runTest.ts — pass err to console.error so the cause and stack trace reach the log.

Verification

• yarn run lint passes (eslint + prettier, including the reformatted YAML).
• yarn run test passes locally.
• Forced a resolution failure against a throwaway copy of the compiled runner to confirm the catch block now surfaces the cause at exactly the point CI went silent:

- Resolving version...
Failed to run tests Error: Invalid version 0.0.0-does-not-exist
    at download (node_modules/@vscode/test-electron/out/download.js:344:23)
    at async downloadAndUnzipVSCode (node_modules/@vscode/test-electron/out/download.js:444:12)
    at async runTests (node_modules/@vscode/test-electron/out/runTest.js:52:40)

Note this PR does not add retry logic for the VS Code download — removing the self-inflicted duplicate run addresses the observed cause, and a real outage should still fail loudly (and now legibly).

🤖 Generated with Claude Code