…rytan#130)

Adds the canonical migration runner framework introduced in upstream's
v0.11.1 PR (Minions + canonical migration + skillify). Takes only the
skeleton — the runner CLI, the registry shell, the shared types, and
the preferences/cli-util support files needed by future orchestrators.
The actual v0.11.0 orchestrator is Minions-adoption-specific (agent
queue setup, AGENTS.md routing, jobs smoke test) and is deliberately
NOT taken — PBrain has not adopted Minions.

Registry starts empty; subsequent commits in this wave add the JSONB
repair orchestrator (v0.12.2) which is the only migration we need
right now. The `compareVersions`, `getMigration`, and migration-list
plumbing are already in place for future additions.

Also wires the matching `pbrain init --migrate-only` branch used by
orchestrators to apply schema migrations against an already-configured
engine without clobbering the user's engine choice.

New commands:
  pbrain apply-migrations [--list] [--dry-run] [--migration vX.Y.Z]

New files:
  src/commands/apply-migrations.ts
  src/commands/migrations/index.ts       (empty registry)
  src/commands/migrations/types.ts
  src/core/cli-util.ts
  src/core/preferences.ts

(cherry picked subset of commit d861336cfa78e4f2ae9b67ecb79da70d0f6bb630)

Co-Authored-By: Garry Tan <garry@garrytan.com>

…an#196)

Data-correctness hotfix for Postgres-backed brains. Pulls forward the
v0.12.2 fix wave from upstream GBrain. PGLite brains were unaffected.

Three related Postgres-string-typed-data bugs that PGLite hid:

1. JSONB double-encode (postgres-engine.ts + files.ts): the
   ${JSON.stringify(value)}::jsonb interpolation pattern made
   postgres.js v3 stringify again on the wire, storing JSONB columns as
   quoted string literals. Every frontmatter->>'key' returned NULL on
   Postgres-backed brains; GIN indexes were inert. Fix: switch to
   sql.json(value), the postgres.js-native JSONB encoder. Affected:
   pages.frontmatter, raw_data.data, ingest_log.pages_updated,
   files.metadata, page_versions.frontmatter.
2. splitBody greedy --- match (markdown.ts): any standalone --- in body
   content was treated as a timeline separator, truncating wiki imports
   by up to 83% (reported by @knee5). Fix: require an explicit sentinel
   (<!-- timeline -->, --- timeline ---, or --- immediately before
   ## Timeline / ## History). Plain --- is a markdown horizontal rule.
3. parseEmbedding NaN scores (utils.ts): Supabase returned embedding
   columns as JSON strings; getEmbeddingsByChunkIds yielded NaN query
   scores. Fix: normalize via parseEmbedding.

Also adds /wiki/ subdirectory type inference (analysis, guides,
hardware, architecture, writing).

New:
- pbrain repair-jsonb [--dry-run] [--json] — standalone repair CLI
- src/commands/migrations/v0_12_2.ts — orchestrator (schema → repair
  → verify → record). Registered in migrations/index.ts.
- scripts/check-jsonb-pattern.sh — CI grep guard against regressions.
  Wired into `bun test` via the check:jsonb npm script.
- test/e2e/postgres-jsonb.test.ts — E2E round-trip regression.

Notes on deferred upstream work: upstream's v0.12.0 orchestrator
(knowledge-graph auto-wire) and v0_12_0.ts are NOT registered in this
fork's migration registry — PBrain has not adopted the knowledge graph
layer. The apply-migrations unit tests have been rewritten to assert
the planner's invariants against v0.12.2 instead of upstream's v0.11.0.

Original fixes contributed by:
- @knee5 (PR garrytan#187 — splitBody, inferType wiki, JSONB triple-fix)
- @leonardsellem (PR garrytan#175 — parseEmbedding, getEmbeddingsByChunkIds)

(cherry picked from commit c0b621923b641eae0e7d6228e50d9cdaa6bd97ae)

Co-Authored-By: Garry Tan <garry@garrytan.com>
Co-Authored-By: knee5 <noreply@github.com>
Co-Authored-By: leonardsellem <noreply@github.com>

…rytan#198)

Adds addLinksBatch / addTimelineEntriesBatch to the BrainEngine
interface. Both engines (PGLite + Postgres) implement the batch as a
single `INSERT ... SELECT FROM unnest(...) JOIN pages ON CONFLICT
DO NOTHING RETURNING 1` statement — 4-5 array-typed bound parameters
regardless of batch size, sidestepping the 65535-parameter cap and
the postgres-js sql(rows, ...) helper's identifier-escape gotcha.

pbrain extract (file-source) now flushes candidates 100 at a time via
the batch API instead of one write per link/entry. On large brains
this drops wall-clock from "tens of minutes" to "seconds" end-to-end.

Schema changes required for ON CONFLICT (from_page_id, to_page_id,
link_type) to match a real unique index:

- links table UNIQUE constraint widened from 2 columns to 3
  (from_page_id, to_page_id, link_type), matching upstream v0.10.3
  knowledge-graph layer. addLink's ON CONFLICT clause updated in
  both engines. Fresh installs get the new constraint from schema.sql
  / pglite-schema.ts; existing installs pick it up via migration v8.
- src/core/migrate.ts picks up upstream's v5-v10 schema migrations.
  v5-v7 (minion_jobs / inbox / attachments) create empty tables that
  this fork doesn't use; they're harmless and keep the schema
  forward-compatible with a future Minions integration. v8-v10 touch
  the links and timeline_entries tables we already have.

Notes on deferred upstream work:
- The DB-source `pbrain extract --source db` path remains deferred
  with the knowledge-graph layer (depends on src/core/link-extraction.ts
  which we haven't taken). File-source extract is what uses the new
  batch API.
- Upstream's v0.12.0 Knowledge Graph auto-wire orchestrator is NOT
  registered (src/commands/migrations/v0_12_0.ts deleted).

(cherry picked from commit 699db50d6cb3b96c3ba3cea8cd55c78f0f9c3bae)

Co-Authored-By: Garry Tan <garry@garrytan.com>

Pulls forward the v0.12.3 reliability fixes — sync deadlock,
search-timeout scoping, tryParseEmbedding for search corruption
tolerance, pbrain orphans command + find_orphans MCP op, and two new
doctor checks (jsonb_integrity, markdown_body_completeness) that
point at the repair-jsonb / sync --force remediation.

What changed:

- src/commands/sync.ts: dropped the outer engine.transaction() wrap
  so importFromContent's per-file transaction isn't nested. PGLite's
  _runExclusiveTransaction is non-reentrant, so the inner call used
  to park on the mutex forever once ~10 files hit the pipeline.
- src/core/postgres-engine.ts: searchKeyword / searchVector now scope
  statement_timeout via sql.begin + SET LOCAL so the GUC can't leak
  onto a pooled connection and clip unrelated long-running queries.
- src/core/utils.ts: new tryParseEmbedding() (returns null + warns
  once per process on bad input) for search/rescore paths where
  availability matters more than strictness. parseEmbedding() stays
  strict for migration/ingest paths.
- src/commands/orphans.ts: new. Domain-grouped report of pages with
  zero inbound wikilinks; --include-pseudo flag, --json, --count.
  Also wired as find_orphans MCP operation.
- src/commands/doctor.ts: +2 reliability checks. jsonb_integrity
  scans pages.frontmatter, raw_data.data, ingest_log.pages_updated,
  files.metadata for jsonb_typeof='string' rows (v0.12.0 residue);
  markdown_body_completeness flags pages with compiled_truth <30%
  of raw source when raw has multiple H2/H3 boundaries.

New tests: test/orphans.test.ts, test/postgres-engine.test.ts,
test/e2e/jsonb-roundtrip.test.ts. doctor.ts and sync.ts existing tests
updated with new check/deadlock assertions.

Upstream's src/core/link-extraction.ts (knowledge-graph layer) is NOT
taken — it's Wave-2 material. All integrated code paths operate on
the already-existing links/timeline_entries schema.

(cherry picked from commit 013b348)

Co-Authored-By: Garry Tan <garry@garrytan.com>