Add opt-in nerve.db retention: message compaction + telemetry pruning

[alex-fedotyev]

Summary

nerve.db grows without bound. On a heavily-used install it reached 767MB, of
which the messages table is 714MB, dominated by the machine-facing blocks
JSON (523MB) and thinking (54MB). That JSON is only needed while a message is
rendered live or being indexed into memU. memU extraction reads content
(gated by the per-session last_memorized_at watermark), and SDK resume
restores context from the .jsonl transcript rather than DB blocks, so old
already-memorized messages can drop blocks/thinking while keeping content
(the UI falls back to content text when blocks is NULL).

This adds an opt-in retention subsystem, disabled by default so merging it
mutates no existing data:

• MaintenanceStore DB mixin:
  • compact_messages nulls blocks/thinking for messages older than
    retention_full_days that are past their session's memorize watermark, in
    a non-starred, non-active session. Keeps content. Idempotent.
  • prune_telemetry / prune_file_snapshots delete append-only rows older
    than retention_days.
  • checkpoint truncates the WAL; vacuum rewrites the file to reclaim
    freed pages.
• RetentionConfig (enabled=False, retention_days=90,
  retention_full_days=30, interval_hours=24).
• A background lifespan task that no-ops unless enabled.
• nerve db prune [--dry-run] and nerve db vacuum commands.
• Reference docs in docs/config.md.

No schema change.

Reclaim model

Nulling columns and deleting rows frees pages to the SQLite freelist (reused by
later writes) but does not shrink the file. PRAGMA wal_checkpoint(TRUNCATE)
truncates only the WAL. Only VACUUM shrinks the file, so it is an explicit
operator step (nerve db vacuum), never on the background loop.

Operator reclaim sequence

1. Set retention.enabled: true (optionally tune the windows) in the local
   config; restart to pick it up.
2. nerve db prune --dry-run to preview.
3. Back up nerve.db, then nerve db prune followed by nerve db vacuum
   (daemon stopped) to shrink the file.

Test plan

• New tests/test_db_retention.py (15 tests): compaction eligibility (old
  • memorized + non-starred + non-active), and every exclusion (recent,
    starred, active, never-memorized, message-newer-than-watermark); content
    always kept; idempotency; dry-run mutates nothing; telemetry/snapshot pruning
    deletes old and keeps new while leaving core tables untouched; combined
    run_retention.
• Full backend suite green except pre-existing, environment-specific
  failures unrelated to this change (docker-mode detection under a container
  env var; codex tests that need untracked local fixtures).
• nerve db prune --dry-run, nerve db prune, and nerve db vacuum
  exercised end-to-end against a throwaway database.
• Config parses the new block, clamps ints to >= 1, defaults to disabled,
  and raises no unknown-key warnings.