fix(playground): saved-replay sidebar + chat resilience (v0.2.3) (#43)

## Summary

Five interlocking playground bugs surfaced during a single user session
— each was masking the next. Fixes them all in one PR; cuts v0.2.3.

- **Saved-replay sidebar ignored manual deletes.**
`vscode.workspace.createFileSystemWatcher` is unreliable for
dot-prefixed paths (notably `.copilotkit/`) on Windows. Replaced with a
Node `fs.watch`-based `FixturesDirWatcher` that retries until the dir
exists. Output channel now logs `[fixtures-watcher] attached`, `fs.watch
<event>`, and `[playground] fixtures dir changed — refreshing sidebar`
so the chain is self-diagnosing.
- **Clicking ▶ on a saved replay produced an empty chat.** The extension
posted `bundle-ready` then `play-fixture` back-to-back; the webview's
`executePlaygroundBundle` is async, so the `play-fixture` `CustomEvent`
fired before the new bundle's `PlaygroundChat` had registered its replay
listener. App shell now unmounts the stale bundle synchronously on
`bundle-ready`, queues the messages in state, and dispatches in a parent
`useEffect` keyed on `[bundle, pendingReplay]` (React runs child effects
before parent effects on the same commit → listener is attached before
dispatch).
- **Clicking ▶ on a deleted fixture permanently bricked the panel.**
`load-fixture` set `replayFixturePath` first, then tried to read. The
read failed but the path stayed, so every subsequent `runBundle` re-read
the missing file and crashed. Refresh did nothing. Now:
read-then-commit; on read failure clear state, refresh the sidebar,
warn. `runBundle` also wraps its fixture read in try/catch and falls
back to record mode when the file vanishes mid-session.
- **Refresh wasn't a recovery path.** `runBundle` now posts a fresh
`fixtures-list` at the *start* of every bundle pass, so Refresh is a
guaranteed manual reconciliation with disk.
- **Chat returned silently with too many `vscode.lm` tools.** Claude
through Copilot Chat returns 200 OK + empty stream when ~80+ tools are
forwarded. `vscode-lm-factory` now auto-retries once without `vscode.lm`
tools, so the user gets a real response on the same query without
changing their settings. If both attempts return empty, an actionable
in-chat message explains the cause and points at
`copilotkit.playground.enableVscodeLmTools`.
- **`LanguageModelDataPart` text/json mimes are now surfaced.** Newer
Claude builds routed through Copilot Chat stream text content as
`DataPart` with `text/plain` or `application/json` mimes — previously
dropped on the floor, producing the same empty-chat symptom. Unknown
part types are logged with their constructor name + mime.

Plus user-driven removal of a stale frontend tool from
`test-workspace/playground/v2/Tools.tsx`.

## Test plan

- [x] `pnpm run check-types` clean
- [x] `pnpm run test` — 312 pass, 1 skipped (added 8 new tests across
`view-provider`, `vscode-lm-factory`, `fixtures-dir-watcher`,
`App.replay-race`)
- [x] Manual: deleted a fixture from disk → sidebar updates within ~1s
- [x] Manual: clicked ▶ on a deleted fixture → toast, sidebar refreshes,
panel does not hang
- [x] Manual: asked "what's the weather in sarajevo" with
`enableVscodeLmTools=true` and 87 vscode.lm tools forwarded → auto-retry
path fires, chat responds normally

fix(publish): drop --target from vsce publish (conflicts with --packa…

…gePath) (#27)

## Why

After #26 surfaced vsce's actual error (which the bash retry function
was previously eating), the diagnostic was clear. Reproduced locally by
downloading the failed run's `vsix-darwin-x64` artifact and running:

```
$ npx @vscode/vsce@^3 publish --target darwin-x64 --packagePath extension-darwin-x64.vsix --pat <fake>

 ERROR  Both options not supported simultaneously: 'packagePath' and 'target'.
        Use 'vsce package --target <target>' to first create a platform specific
        package, then use 'vsce publish --packagePath <path>' to publish it.
```

vsce reads the target from the `.vsix` manifest's `<TargetPlatform>` tag
(stamped at `vsce package --target` time) and explicitly rejects
`--target` together with `--packagePath`. The error fires before any
auth check, which is why all 4 vsce publishes failed in ~1 second flat.

ovsx is more permissive and accepts both flags together — that's why
every ovsx publish succeeded while every vsce publish failed
identically.

## Fix

Drop `--target ${{ matrix.target }}` from the `vsce publish` call. The
matrix still records the target via the step name, the artifact name,
the .vsix filename, and the manifest stamping — nothing self-documenting
is lost.

`ovsx publish --target ... --packagePath ...` is left as-is since ovsx
accepts both and the explicit `--target` round-trips through the ovsx
API for the registry's own metadata (and worked on the previous run for
all 4 targets).

## State going in

- **VS Code Marketplace:** still on `0.2.0`. `0.2.2` never landed.
- **Open VSX:** has `0.2.2` for all 4 targets.
- **#26 merged:** retry function now surfaces vsce/ovsx output
regardless of exit code.

## Recovery on merge

`prepare` will see Marketplace doesn't have `0.2.2` → proceeds. ovsx
will return "already exists" for all 4 targets → idempotent success.
vsce, with the conflicting flag removed, should publish all 4 targets
cleanly. `finalize` will tag and cut the GitHub Release.

## Test plan

- [x] Reproduced the error locally with the failed-run artifact
- [x] Confirmed `vsce publish --packagePath <vsix>` (without `--target`)
is the documented vsce form for publishing pre-packaged
platform-specific .vsix files
- [ ] After merge: push to main triggers re-run; vsce publishes all 4
targets to Marketplace; finalize creates `vscode-extension-v0.2.2` tag
and GitHub Release

Playground chat tab: vscode.lm, scanner, bundler, fixture replay (#22)

## Summary

Migrated from CopilotKit/CopilotKit#4635 (by @AlemTuzlak) to the
standalone repo.

- Replaces the OpenAI/Anthropic + API-key + Node-subprocess flow with
VSCode's built-in Language Model API (`vscode.lm`), with optional
`vscode.lm.tools` forwarding so the model can use Copilot's web-fetch
tool out of the box
- Adds a playground scanner that detects `<CopilotKit>` JSX nodes,
serializes props, walks ancestors, maps hooks to components, and
orchestrates multi-file scanning
- Codegen modules (aggregator, entry, error-boundary, provider-chain
templates) and rolldown bundler with Tailwind v4 compile pass + file
watcher for hot-reload
- Custom `PlaygroundChat` that drives the runtime over SSE, executes
frontend tool handlers locally, supports `followUp: false`, renders
markdown, and animates fixture replay
- Fixture store for `.copilotkit/fixtures` with v2 schema
(`RecordedCall[]`) and replay mode
- Resizable/collapsible panels: MountedComponents, ConversationSidebar,
DiagnosticsPanel, ModelPicker, ScannerView (state persisted to
localStorage)
- Test-workspace rebuilt as a v2-only weather demo with
`useFrontendTool` and deterministic mock-data handlers

## Test plan

- [ ] `vitest run` passes locally (300 tests)
- [ ] Build produces clean dist
- [ ] Open dev host on `test-workspace`, open Chat tab, ask "what's the
weather in Sarajevo?" — renders populated WeatherCard