Fix ingester DB cache eviction/export safety and startup cleanup

[ian-ross]

Summary

• fix staged-connection eviction so touched (dirty) staged DBs are committed before eviction
• fail fast when opening staging DBs for VACUUM export if read-only open fails
• close nursery DB explicitly after promotion in connect-path nursery eviction
• narrow startup scratch cleanup to only exporter temp files
• add regression tests for touched-connection eviction commit safety and strict staging-open failure

Motivation

PR #9 had post-review correctness gaps around data-loss risk and fail-open behavior during staging export/import. This PR addresses those before merging the ingester isolation work.

Changes

P1: Prevent uncommitted touched DB evictions

_evict_staged_connections now loops while over cache limit and for touched candidates:

• tries to commit pending DBs and retries eviction
• raises if commit does not clear touched state

P2: Fail-fast staging export connection behavior

_open_raw_staging_connection now raises RuntimeError on read-only open failure instead of falling back to read-write.

P3: Resource/cleanup hardening

• connect-path nursery promotion now explicitly closes promoted nursery DB after promotion
• startup cleanup in scratch/ingester-export now removes only *.export.*.sqlite files, preserving unrelated scratch artifacts

Tests

• uv run --python 3.13 pytest tests/feder_ingest/test_db_cache_paths.py -q