feat(chat): Prompt Explorer modal for iterative prompt enrichment

[vibegui]

Summary

Adds an Explore button to the chat composer that opens a modal for turning a rough idea into a richer, more complete prompt. A fast-tier model rewrites the draft in the user's own voice and language, inserts [bracketed] fill-in placeholders for missing details, and grows the prompt gradually (~2-3×/iteration) rather than ballooning it.

• Backend: new lightweight SSE route POST /api/:org/prompt-explorer/stream (apps/mesh/src/api/routes/prompt-explorer.ts) — calls the ai SDK streamText against the org's fast tier (mirrors suggest-commit-message.ts); no thread/NATS machinery. Registered in org-scoped.ts.
• System prompt: pure, unit-tested builder (apps/mesh/src/lib/prompt-explorer-system.ts) that preserves voice/grammatical-person/language and enforces a length budget.
• Frontend: usePromptEnricher hook (use-prompt-enricher.ts) + PromptExplorerDialog (prompt-explorer-dialog.tsx); versions sidebar with Iterate + rollback; Enter → Iterate, Cmd/Ctrl+Enter → Send, Shift+Enter → newline. Wired into chat/input.tsx.

How to test

1. Open a chat in an org that has an AI provider configured.
2. Click the ✨ Explore button in the composer (right side, near the model/send buttons). If the composer already has text, it auto-runs the first enrichment.
3. Type a rough idea (≥10 chars) → the right panel streams an enriched, first-person, same-language version with [blanks].
4. Iterate snapshots the current version into the sidebar and refines from it; click a previous version to roll back.
5. Fill in blanks, then Send (or Cmd/Ctrl+Enter) — the optimized prompt is injected into the composer and sent.

Notes / context for reviewers

• This PR is partly to get a clean-environment test: locally, the streamed finish frame + connection-close were intermittently withheld by the dev tunnel (*.localhost), which made the spinner appear to hang forever even though the server completes in ~1.4–2.7s. Verified server-side (direct + Vite + :80) that the stream delivers finish and closes cleanly, and that output is first-person + capped. The idle-completion guard in usePromptEnricher (resolve after ~2s of token silence) makes the UI robust to that regardless of transport. Curious whether others see the spinner resolve cleanly without the tunnel quirk.
• bun run fmt, bun run lint (no new warnings), bun run check, and the prompt-explorer-system unit tests all pass.

🤖 Generated with Claude Code

---

Summary by cubic

Adds an Improve prompt modal in chat that turns rough ideas into a concrete, ready-to-use prompt in the user’s own voice and language, expanding gradually each iteration. Results stream from the org’s fast tier and can be improved again or sent to the composer.

• New Features
  • "Improve" button in the composer opens a modal with versions, Improve, and rollback; Enter → Improve, Cmd/Ctrl+Enter → Send.
  • Lightweight SSE route POST /api/:org/prompt-explorer/stream streams enriched text and reasoning from the org’s fast tier.
  • Pure system‑prompt builder that preserves voice/person/language, avoids placeholders, and enforces a gradual length budget; unit tests included.
  • usePromptEnricher hook streams text/reasoning with cancel support; status is driven by useState for reliable completion on finish or connection close.

^{Written for commit 8d4ab28. Summary will update on new commits.}