Persist Brain corpus for nightly synthesis

[mrubens]

What changed

• persist the Brain's Markdown corpus under /data/brain
• configure gbrain's built-in synthesis and pattern phases to use that corpus
• keep Postgres as the search index, extracted-fact store, and durable job backend
• perform a clean one-time cutover by recreating only the isolated gbrain database and corpus, then relying on Roomote's existing client reprovisioning and collector backfill
• install Git in the gbrain image for its filesystem maintenance phases
• update public and Coolify operator documentation

Why

The Postgres-backed Brain had no configured filesystem corpus. Nightly maintenance could run database-only phases, but gbrain skipped synthesis and other filesystem phases because it had no Brain directory to read or write.

This gives gbrain the on-disk corpus its built-in maintenance expects without adding another database or a Roomote-specific synthesis implementation.

Impact

The first boot with this storage layout intentionally reinitializes an existing Postgres-only Brain. Roomote detects the missing scoped clients, resets its ingestion checkpoints, and repopulates completed tasks and connected integration sources. Later restarts preserve the corpus and database.

Validation

• sh -n .docker/gbrain/entrypoint.sh
• built the gbrain Docker image
• booted it against a real Postgres/pgvector container and verified the layout initializes once across restarts
• wrote a page through gbrain and verified the Markdown file appeared under /data/brain
• mise exec -- pnpm --filter @roomote/docs check
• mise exec -- pnpm deployment:validate
• git diff --check

[roomote-community]

No new code issues found. See task

• Cutover can repeat after interrupted initialization (.docker/gbrain/entrypoint.sh:74).

_{Reviewed 8ede28f}