feat: consume @supabase/postgrest-typegen for type generation

CLAUDE.md

@@ -76,7 +76,7 @@ npm run format          # Format code with Prettier
    - `app.ts`: Main Fastify app with routes, CORS, Swagger docs
    - `admin-app.ts`: Admin server (runs on PG_META_PORT + 1) for metrics
    - `routes/*.ts`: REST endpoints mapping to library methods
-   - `templates/*.ts`: Type generation templates for different languages
+   - `routes/generators/*.ts`: Type generation endpoints backed by `@supabase/postgrest-typegen`
 
 ### Object Manager Pattern
 
@@ -116,9 +116,9 @@ Database errors are formatted to mimic `psql` output:
 
 Type generation (`npm run gen:types:*`) works by:
 1. Connecting to a database (test DB or custom via `PG_META_DB_URL`)
-2. Fetching all schemas, tables, columns, relationships, functions, types
-3. Passing data to language-specific templates in `src/server/templates/*.ts`
-4. Templates output type definitions to stdout
+2. Introspecting schemas, tables, columns, relationships, functions, and types via `@supabase/postgrest-typegen`'s `introspect()` (wrapped by `src/lib/generators.ts`)
+3. Passing the metadata to the package's language generators (`generateTypescript`, `generateGo`, `generateSwift`, `generatePython`)
+4. Generators output type definitions to stdout
 
 Environment variables:
 - `PG_META_GENERATE_TYPES`: Language (typescript, python, go, swift)
@@ -154,20 +154,27 @@ PG_META_MAX_BODY_LIMIT_MB=3             # Max request body size in MB (default:
 PG_META_SHUTDOWN_GRACE_PERIOD_SECS=10   # Shutdown wait on in-flight work before force-exit (default: 10)
 ```
 
-Type generation formatting (opt-in):
+TypeScript generation on a worker thread (opt-in):
 ```bash
-PG_META_FORMAT_IN_WORKER=true           # Format generated types on a worker thread (default: false)
-PG_META_FORMAT_POOL_SIZE=1              # Worker threads used for formatting (default: 1)
-PG_META_FORMAT_MAX_QUEUE=20             # Max format calls in flight before returning 503 (default: 20)
-PG_META_FORMAT_TIMEOUT_SECS=60          # Per-format timeout (default: 60)
+PG_META_FORMAT_IN_WORKER=true           # Generate TypeScript types on a worker thread (default: false)
+PG_META_FORMAT_POOL_SIZE=1              # Worker threads used for generation (default: 1)
+PG_META_FORMAT_MAX_QUEUE=20             # Max generation calls in flight before returning 503 (default: 20)
+PG_META_FORMAT_TIMEOUT_SECS=60          # Per-generation timeout (default: 60)
 PG_META_FORMAT_IDLE_TIMEOUT_SECS=30     # Idle time before a worker exits (default: 30)
 ```
 
-Prettier is CPU-bound and synchronous, so formatting a large schema blocks the
-event loop for seconds and the server cannot answer anything else, health checks
-included. `PG_META_FORMAT_IN_WORKER=true` moves it to a worker thread. Always
-off during type generation (`PG_META_GENERATE_TYPES`), which generates once and
-exits.
+`generateTypescript` is CPU-bound and synchronous: on a large schema it blocks
+the event loop and the server cannot answer anything else, health checks
+included. Formatting is the bulk of it (oxfmt since postgrest-typegen 0.2.0,
+much faster than the prettier it replaced but still synchronous) and the string
+building ahead of it costs too, so `PG_META_FORMAT_IN_WORKER=true` moves the
+whole call to a worker thread (`src/server/format-pool.ts`). Metadata
+crosses the thread boundary as a structured clone, which is plain JSON and does
+not measurably change wall-clock time. Always off during type generation
+(`PG_META_GENERATE_TYPES`), which generates once and exits.
+
+The env vars keep their `PG_META_FORMAT_*` names from when the worker formatted
+only, so existing deployments do not need reconfiguring.
 
 ## Testing Notes