ci: run extension host tests on Windows in a merge-only job - #16619

Moves Windows extension host tests into their own merge-only workflow job, and fixes what that turned up on Windows. Fixes #8853.

Today the only Windows ext-host coverage is positron-r, run on shard 1 of the Windows e2e job (added in #8816). That step is soft-fail: continue-on-error, always exit 0. Its Slack alert in test-merge.yml reads needs.e2e-windows-electron.outputs.extensions_failed, but #11014 renamed that job to e2e-windows, so the alert has never fired since. Meanwhile positron-r is failing 4 tests on Windows on main (e.g. merge run 38429410345, shard 1: 204 passing, 4 failing).

Workflow changes

• New test-ext-host-windows.yml, called from test-merge.yml and test-full-suite.yml (new run_ext_host_tests_windows input). It is not called from the PR workflow. It runs the Electron driver via scripts\test-integration.bat and fails the job when tests fail.
  • The web (Chromium) driver is not run: Positron has no web build on Windows.
  • The Remote driver is not run: Windows remote (WSL) is covered by e2e tests.
  • Restores the Windows build caches but does not save them; the e2e-windows job in the same merge run owns the save.
  • Sets VSCODE_SKIP_PRELAUNCH=1 on the test step, as upstream's Windows CI does. Prelaunch already ran, and otherwise the agent host E2E runner calls spawnSync('npm.cmd') without shell: true, which Node rejects with EINVAL.
• test-e2e-windows-run.yml: removes the positron-r step, its log save/upload steps, and the extensions_failed output. Removes the dead slack-notify-win-extensions job from test-merge.yml. PRs tagged @:win no longer run positron-r ext-host tests on Windows; that coverage moves to merge.
• Insights filter_jobs regex in both callers now matches ext-host-win.
• test-full-suite.yml is now at the 25-input limit for workflow_dispatch.

scripts/test-integration.bat

• Adds the 13 Positron extension suites that test-integration.sh runs, and skips colorize as the .sh does (Test: fix MS colorize integration tests #7349). The Positron suites are skipped when --suite is given, since they are not selectable by name.
• Resolves TMP/TEMP to the long path, the same line upstream already has in test-remote-integration.bat and test-web-integration.bat. On CI TMP is C:\Users\RUNNER~1\..., and under Node 24.16+ fs.watch on an 8.3 short-name path drops every event (Potential regression with libuv in Node 24.16.0 nodejs/node#63638; plain Node asserts in libuv fs-event.c).
• Runs the Positron suites with --config pointing at .vscode-test.js through a lowercase drive letter. The extension host loads an extension through c:\, while vscode-test required the test files through C:\. Node keys its module cache by path, so a test importing extension modules got a second, never-initialized copy (confirmed with a probe: both require.cache keys present while the extension was active). This is what broke positron-code-cells' context and decoration tests and positron-r's RSessionManager.

Product bugs found (Windows only)

• next-edit-suggestions: exact-filename keys in nextEditSuggestions.enabled (e.g. "Dockerfile": false) never matched on Windows, because the basename was taken from document.fileName at the last / only. It now comes from path.posix.basename(document.uri.path), as isFileExcludedFromAI in the same file already does, with a Windows-path test so Linux CI covers it.
• positron-data-driver-duckdb: on Windows DuckDB reports a lock conflict as a sharing violation (... File is already open in <process> (PID n)), so users got the raw error instead of the "another session has it locked" explanation. The pattern now also matches DuckDB's own File is already open in text (the OS part of the message is localized).

Test fixes

• positron-r: rootSignature builds expectations with path.join (faking process.platform does not change path); session-dispose stubs RSessionManager.instance instead of depending on activation order, which also keeps its fake sessions out of the live manager.
• authentication: snowflake path expectation uses path.join. Tests that start the provider catalog watcher remove their temp dir with an async fs.promises.rm retry: Windows releases a closed watcher's directory handle on a later event loop turn, and rmSync's own retries block the loop.
• positron-data-driver-odbc: resolveUnixConfigPaths uses path.posix. The module takes the platform from an injected host and already uses path.posix/path.win32 explicitly elsewhere; this was the one spot using the host path. No change on Unix.
• CSS / HTML language server tests (upstream files, Positron markers): their entry points globbed path.posix.join(testRoot, ...); on Windows testRoot has backslashes, which glob reads as escapes, so both suites ran 0 tests and passed. Now glob relative to testRoot, and fail if no test files match. Windows runs the same 9 CSS / 71 HTML tests as Linux. Upstream has the same code.
• positron-data-driver-sqlite / duckdb: teardowns remove temp dirs with an async retry. disconnect() kills the worker process without waiting for it to exit, so the database file is briefly still open, and Windows will not delete an open file. (Making disconnect() await worker exit would be the product-side fix; left out of scope here.)

Unrelated

• configureSnippets.ts: one-line indent fix from the formatter. The file is unformatted on main since fix: snippet picker matching after upstream label/description swap #16607, and the pre-commit hook blocked the merge commit that synced this branch with main.

Release Notes

New Features

• N/A

Bug Fixes

• Next Edit Suggestions: per-filename entries in nextEditSuggestions.enabled now apply on Windows.
• DuckDB data connections: opening a database another session has locked now explains the conflict on Windows too.

Validation Steps

• Locally on Windows, all 13 Positron suites run the way the .bat now runs them: 12 pass. positron-r's R-session tests could not start an R kernel from this machine's agent-launched shell (a known local environment issue with Ark's ZeroMQ poller); main's Windows e2e job shows those tests starting R sessions fine on the CI runner.
• The originally failing tests were reproduced locally first and fail without the fixes.
• Full suite runs on this branch with only run_ext_host_tests_windows:
  • Run 1 (38450231603): failed before any tests, on the agent host E2E npm.cmd EINVAL.
  • Run 2 (38455222586): node.js integration, agent host E2E, and every upstream ext-host suite passed on Windows; stopped at authentication.
  • Run 3 (38466440126): green. Every suite passed, including positron-r (208 passing, with R sessions on the CI runner). CSS/HTML reported 0 tests, which led to the fix above.
  • Run 4 (38473148320): green on 05908324ee, now including 9 CSS / 71 HTML server tests. Review follow-ups after it (7dcca78d53: empty-glob guard, NES basename, job timeout 150, no id-token) were verified locally.
  • Run 5 (38478932934): green on 7dcca78d53 (the PR head). All 33 sections of test-integration.bat ran with no failures; the job took about 1 hour.