# SQLite startup fix

## Problem

Some installations aborted during `graph.Open` with an opaque modernc SQLite error similar to:

```text
SQL logic error: out of memory (1)
```

The original startup path applied a 64 MiB SQLite page cache, a 256 MiB mmap window and `temp_store=MEMORY` while the connection was being opened. Those settings are unnecessary because the live graph and embeddings already reside in Go memory. The DSN-based PRAGMA list also made it impossible to identify which startup step failed.

## Changes

- Open SQLite using a plain absolute filename.
- Verify that `BRAIN_DATA_DIR` exists, is a directory and is writable before opening SQLite.
- Reject a `graph.db` path that is accidentally a directory.
- Apply PRAGMAs individually with step-specific error messages.
- Use conservative storage defaults:
  - `cache_size=-8192` (approximately 8 MiB)
  - `mmap_size=0`
  - `temp_store=FILE`
  - WAL, `synchronous=NORMAL`, foreign keys and a 5-second busy timeout remain enabled.
- Persist and report the journal mode actually selected by SQLite.
- Add the dependency checksums to `go.sum` and copy both `go.mod` and `go.sum` in the Docker dependency layer.

## Recovery

Because this is a staging-only database, remove files left by a failed first start before rebuilding:

```bash
docker compose stop brain
docker compose run --rm --no-deps --entrypoint sh brain -c '
  rm -f /app/data/graph.db /app/data/graph.db-wal /app/data/graph.db-shm
'
docker compose up -d --build --force-recreate brain
```

If startup still fails, the new error includes the exact phase, for example an unwritable data directory or a failed WAL configuration.