fix: disable llama embeddings by default, switch to opt-in via BRAIN_EMBED_ENABLE #115

[four-bytes-robby]

Changes

• Flip BRAIN_EMBED_DISABLE → BRAIN_EMBED_ENABLE (opt-in, default off)
• Update @four-bytes/opencode-plugin-lib 0.2.0 → 0.3.0 (EventBus export)
• Bump version 1.7.0 → 1.7.1

Closes #115

---

Summary by cubic

Switches llama embeddings to opt-in (off by default) and improves DB resilience and responsiveness so search isn’t blocked during ingest. Addresses #115.

• New Features

  • Real embeddings are opt-in via BRAIN_EMBED_ENABLE; graceful fallback to hash-based vectors when disabled or unavailable.
  • EmbeddingService adds a priority queue (search ahead of ingest), batch embeds, and a Promise-locked init.
  • TUI shows a “connecting…” state with stale detection; logs are streamed via app.log for clearer diagnostics.
  • Added scripts/test-llama.ts to validate node-llama-cpp binaries and model loading.

• Refactors

  • SQLite hardening: busy_timeout 30s, WAL checkpoint at startup and after large ingests, and withDbRetry() for SQLITE_BUSY.
  • Ingest and chunking yield to the event loop (between files and every 50 chunks) to keep the UI and tools responsive.
  • Search/memory ops wrapped with retries and better timeout handling; clearer user-facing messages on contention.
  • Embedding pipeline uses session cache and batch insert; reduced console noise unless BRAIN_DEBUG is set.

• Migration

  • To enable real embeddings, set BRAIN_EMBED_ENABLE=true (BRAIN_EMBED_DISABLE is no longer used).
  • After upgrading, run a fresh install to pick up the postinstall fix that symlinks node-llama-cpp bins (Linux x64).

^{Written for commit 681edd1. Summary will update on new commits.}