feat(vscode): skill-backed @ctx participant, verified against the real CLI

[bilersan]

feat(vscode): skill-backed @ctx participant, verified against the real CLI

Closes #127. Reworks #128 (closed for rework) onto current main, addressing every
point of that review. Opened as a draft because the skill-command design changed
(see Design questions) and I'd like your read on it before a full review.

Spec: specs/vscode-skill-backed-participant.md (committed on-branch; it contains the
full #128 review-point table).

Why now

main's @ctx participant is currently dead: every call passes --no-color, which the
CLI no longer accepts (30 call sites), so each command prints cobra usage instead of
output. package.json also declares /hook and /diag, which nothing dispatches, while
/setup is dispatched but not declared.

What ships

• 27 CLI-backed commands re-targeted to the current tree: <noun> add with
  --session-id/--branch/--commit provenance, journal source, hook notify|message|pause|resume,
  index, sysinfo, usage, task complete|archive|snapshot, …
• 9 skill-backed commands (/brainstorm /spec /implement /next /remember /reflect /wrap-up /blog /consolidate) send the canonical ctx-<name> SKILL.md (bundled at build
  time) to the chat model, grounded in ctx agent output and #file attachments. The model
  proposes commands; it cannot run them and never claims to.
• runCtx: no shell (fixes Windows argument splitting and & command injection), stdin
  closed, exit code surfaced; non-zero exits render as failures, never as results; cancel,
  timeout, spawn failure and output overflow are errors.
• Parity guard (the core of the feat(vscode): skill-backed chat participant (45 commands) #128 review): commandParity.test.ts drives every command
  branch through the handler and snapshots each ctx argv to src/ctx-cli-surface.json;
  internal/bootstrap/vscode_surface_test.go validates every argv against the cobra tree,
  actually runs the entry add calls in a scratch project, and checks each referenced skill
  ships — so a CLI rename now fails go test. Seeded with 14 of feat(vscode): skill-backed chat participant (45 commands) #128's dead calls and 2
  missing skills, it catches all of them. Refresh: npx vitest run -u in editors/vscode,
  review the diff, rerun the Go test.
• Reminder bell reads ctx remind list (and no longer refreshes before the binary resolves);
  /pause /resume → hook pause|resume; pre-init gate removed (the CLI decides); commands
  use the active editor's workspace folder in multi-root windows.
• Removed: /prompt /dep /reindex (CLI gone), /loop, /site. Not carried over from
  feat(vscode): skill-backed chat participant (45 commands) #128: /verify /audit (no shipped skill), /map /check-links /changelog /worktree
  (need repo browsing / git the participant can't do; ctx-architecture also still references
  the removed ctx deps). Guardrails / terminal capture from feat(vscode): skill-backed chat participant (45 commands) #128 are not included.
• Version 0.10.0 (CHANGELOG "0.10.0 - Unreleased"); the 0.9.0 section is byte-identical.

Verification

• editors/vscode: npm ci, build, tsc (default + tsconfig.ci.json), eslint,
  vitest 82/82, vsce package — all green.
• Go: new tests pass; full go test ./... failure roster identical before/after on Windows
  (pre-existing Windows-only failures), go vet clean, golangci-lint v2.13.2 0 issues.
• Every recorded argv executed against a fresh ctx build in a scratch project.
• Manually tested the packaged .vsix in VS Code 1.139.1 and VS Code Insiders 1.140.0
  on Windows against a scratch ctx init project.

Design questions for you

1. Skill-backed commands: OK to send the real SKILL.md to the chat model (read-only,
   proposes commands) instead of feat(vscode): skill-backed chat participant (45 commands) #128's TypeScript re-implementations?
2. Dropped wishlist commands (/verify /audit /map /check-links /changelog
/worktree): drop, or keep as follow-ups once the participant can use tools?
3. Version: hack/release.sh rewrites the extension version to ctx's VERSION at release,
   which would move 0.10.0 back to 0.9.x — keep 0.10.0 "Unreleased" or let the release sync it?
4. /add task has no --section default because the CLI refuses a catch-all section —
   fine, or should the participant pick one?
5. Session ID: VS Code exposes no AI session id, so --session-id uses the first 8 chars
   of vscode.env.sessionId. Acceptable?
6. Test location: the Go parity test reads editors/vscode/src/ctx-cli-surface.json —
   fine in internal/bootstrap, or would you prefer it elsewhere?

Noticed while reconciling (not changed here): task add / convention add help and skills
call --section optional but the CLI requires it; pad edit's doc comment says it opens
$EDITOR, the code doesn't.

🤖 Generated with Claude Code