fix(kanban): re-probe DB health on initialized paths

[jamesraddock]

Summary

• separate process-lifetime schema initialization from bounded DB-health caching
• re-run the integrity guard after a 30-second healthy TTL on the initialized-path fast path
• keep steady-state connects lock-free inside the TTL, then use the existing bounded init flock so fix(kanban): auto-REINDEX index corruption, backup retention cap, periodic WAL checkpoint, repair CLI #68654's narrow REINDEX repair remains serialized
• invalidate health stamps when a board is removed, repaired, or explicitly reinitialized

Context

#41795 was superseded by #68654 for backup retention and index-only repair, but its TTL re-probe did not land. Without this follow-up, a long-lived gateway dispatcher can keep trusting its first-open health verdict for the lifetime of the process.

This PR ports only that remaining TTL behavior onto the current fast path. It does not revive the superseded guard stack or change wal_autocheckpoint.

Testing

• Rebased onto current main (781968be) containing fix(state): refuse WAL on SQLite builds with the WAL-reset bug #70055.
• Unpatched-current-main behavioral regression: fails — even after the new SQLite 3.50.4 gate selects journal_mode=DELETE, the initialized fast path never reaches the health guard.
• Same focused regression with this commit: passes.
• tests/test_sqlite_wal_reset_gate.py: 23 passed.
• Combined Kanban DB/init/repair/board + WAL-gate run: 340 passed, 2 failed. Both failures reproduce on unpatched current main and are fix(state): refuse WAL on SQLite builds with the WAL-reset bug #70055 test-assumption regressions, not this diff:
  • test_wal_checkpoint_truncates_wal_file assumes a WAL sidecar exists, while the new vulnerable-runtime gate deliberately selects DELETE.
  • test_connect_falls_back_to_delete_on_locking_protocol expects a warning already consumed by fix(state): refuse WAL on SQLite builds with the WAL-reset bug #70055's process-global warn-once state when the full file runs.
• Ruff on both changed files — passed

[teknium1]

Thanks for the focused follow-up. The premise remains present on current main: hermes_cli/kanban_db.py:1940 makes the integrity guard return for initialized paths, while the initialized fast path at hermes_cli/kanban_db.py:2135 opens and returns a connection without invoking that guard.

PR b69ba22f78864a4751de07014ada7b321c6d8193 separates health freshness from schema initialization, uses the existing bounded init lock for the expiry path, and adds a regression covering the cheap-within-TTL path, the re-probe, and fail-closed behavior after corruption. No correctness or design-fit issue was identified from the reviewed diff.

Automated hermes-sweeper review.