Audited tobi/qmd's 458 commits since our fork divergence, filtered to
post-cleanup-compatible fixes (skipped node-llama-cpp / Qwen3 / GGUF
specific ones since we removed those layers in the 2026-04-13 cleanup).

Already had (verified during audit, no action):
- 09a4d19 fix(store): error on embedding dimension mismatch (#501)
- ef062e1 fix(multi-get): brace expansion glob (#424)
- 0212363 fix(mcp): read version from package.json (#431)

Cherry-picked tonight:
- 77e71d0 fix: USERPROFILE fallback for Windows HOME (max0n232)
  src/store/path.ts: HOME = HOME || USERPROFILE || /tmp. Critical for
  MCP subprocess case on Windows where Claude Code passes USERPROFILE
  but not HOME, causing QMD to open an empty /tmp DB.

- 9dd8a73 fix(mcp): enableProductionMode before getDefaultDbPath (jungholee)
  src/mcp/server.ts: imports + calls enableProductionMode at module
  init. CLI users already get this via cli/qmd.ts:110, but SDK
  consumers importing the MCP server directly hit the production-mode
  guard on getDefaultDbPath. This is the same race upstream fixed.

- 0adbdeb fix(store): surface actionable sqlite-vec guidance (Tobi)
  src/store/db-init.ts: track _sqliteVecUnavailableReason at init
  time so ensureVecTableInternal can throw with the original cause
  instead of a generic "extension not available" message. Reuses our
  existing createSqliteVecUnavailableError helper.

- 17074ea fix: include line in CLI --json output (rymalia, #533)
  src/cli/qmd.ts: outputResults JSON branch now emits snippetInfo.line
  alongside the snippet so JSON consumers can navigate to the right
  source line.

Skipped (post-cleanup incompatible):
- f53ee26 fix: detect non-GGUF model files - we removed node-llama-cpp
- 6db34d7 fix(llm): catch GPU init failures - llama.cpp specific
- 8644fa9 fix(store): thread embed model URI to format functions -
  uses Qwen3 detection (formatQueryForEmbedding Qwen3 branch) which
  we ripped out tonight. The format function exists, but the Qwen3
  conditional doesn't.
- 54550a3 fix(llm): explicit embed context size - llama.cpp specific
- 1ecb5c9 Fix QMD_LLAMA_GPU backend override - QMD_LLAMA_* env vars
  removed in cleanup
- 26e3d0c fix(status): avoid build attempts during device probe -
  device probe was LlamaCpp-specific, removed in cleanup
- handelize chain (9fb9de4 -> 9c9de94 -> 828823d -> fee576b) - upstream
  briefly removed .toLowerCase(), broke things, reverted, added a
  migration for users hit by the broken window. We never participated
  in that experiment so the chain is a no-op for us.

Deferred (real feature, too invasive for tonight):
- 8404cc3 fix(uri): include index in custom qmd links - adds index
  segment to qmd:// URI parsing so CLI commands can switch indexes
  inline. Requires schema change to VirtualPath type (currently
  {collectionName, path}, would need optional indexName field) plus
  call sites in get/multi-get. Worth porting in a separate session.

Also need to verify (separate work):
- 3023ab3 fix: bump transitive deps for security alerts - need to
  cross-reference against our package-lock.json since our dep set
  diverged after the node-llama-cpp removal.

Tests: pick-vector-matches + memory-eviction + mcp suite all green
(66 pass / 3 skipped). Full suite deferred until phase 5 completes.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

CLAUDE.local.md version bump v2.0.1 → v2.1.0. Two session summaries:
upstream rebase sync (45 commits integrated via rebase) and v2.1.0
testing session (PR tobi#533 submitted, bugs found in JSON line field
and multi-collection search, comments on tobi#383, tobi#463, tobi#217, tobi#241).

CLAUDE.local.md version bump v2.0.1 → v2.1.0. Two session summaries:
upstream rebase sync (45 commits integrated via rebase) and v2.1.0
testing session (PR tobi#533 submitted, bugs found in JSON line field
and multi-collection search, comments on tobi#383, tobi#463, tobi#217, tobi#241).