fix(training-agent): defer pickTaskRegistry pool lookup so prod uses Postgres on cold boot

[bokelley]

Summary

The diagnostic logging from #4067 + the state-store fix in #4072 surfaced a long-standing silent regression:

"Database not initialized" at pickTaskRegistry (registry.js:124)

Same root cause as #4072. mountTenantRoutes() runs before initializeDatabase(), so getPool() at construction throws. pickTaskRegistry's try/catch silently downgrades to InMemoryTaskRegistry — meaning every cold-booted production machine has been running with the in-memory registry since #463 shipped, defeating the whole point of adcp_decisioning_tasks. Buyer creates a media buy on machine A, polls on machine B, sees "task not found" with ~50% probability.

What changed

Wrap the pool in the same lazy PgQueryable adapter pickStateStore uses (#4072). getPool() runs at first query, not at construction; by then the DB is initialized.

const lazyPool = {
  query: (text: string, values?: unknown[]) => getPool().query(text, values),
};
return createPostgresTaskRegistry({ pool: lazyPool });

Drops the try/catch entirely — there's no scenario where construction can fail now, and a real DB error at first query should surface, not be swallowed.

Test plan

• npx tsc -p server/tsconfig.json --noEmit clean
• tenant-smoke 3/3 pass
• After merge: deploy completes; fly logs show Tenant registry initialized with no preceding Database not initialized from pickTaskRegistry

🤖 Generated with Claude Code