fix(dashboard): probe last_read_at in read-only session-store staleness check

[wangyunyou]

Summary

The read-only session-store staleness probe (_SESSION_DB_READ_PROBE_SQL in hermes_cli/web_server.py) was not updated when last_read_at landed in ec0c8d9c2 ("feat(state): sessions carry read/unread state").

Read-only opens skip _reconcile_columns() (by design), so an older store missing only that column passes the probe — archived / pinned / active / compacted all exist — and never triggers the heal-once writable open. Every GET /api/sessions poll then 500s with no such column: s.last_read_at until something opens the store writable (CLI/gateway startup). Profiles used only through the desktop/dashboard backend never get that writable open, so their session lists stay broken indefinitely.

Reproduction

On a profile whose state.db predates ec0c8d9c2 (sessions table without last_read_at):

1. Desktop sidebar polls GET /api/sessions → 500 on every poll (once per ~60s)
2. Logs: hermes_cli.web_server: GET /api/sessions failed → sqlite3.OperationalError: no such column: s.last_read_at (traceback in list_sessions_rich, hermes_state.py:6055)
3. The stale-schema heal in _open_session_db_for_profile never fires because the probe SQL doesn't reference the missing column, so the store stays read-only forever

Observed on a real install: two profiles (xuanru, money) created Aug 2 with stores missing last_read_at; both session lists 500'd from the moment the update landed until a CLI run opened the store writable.

Fix

Add (SELECT last_read_at FROM sessions LIMIT 1) to the probe. A store missing the column now fails the probe, which triggers the existing one-time writable open; _init_schema() → _reconcile_columns() then adds the column declaratively (no version-gated migration needed).

Test Plan

• scripts/run_tests.sh tests/hermes_cli/test_web_server.py tests/test_web_server_sessiondb_eventloop.py -q → 142 passed, 0 failed
• E2E against a real pre-update state.db snapshot (missing last_read_at):
  • probe now raises no such column: last_read_at ✅
  • writable SessionDB open adds the column (53 cols) ✅
  • re-probe passes, list_sessions_rich(limit=3) returns rows ✅

[alt-glitch]

This was generated by AI during triage.

Duplicate of #80047. Both patches add last_read_at to the same read-only session-store probe; #80047 is the earlier open fix and also includes the regression test.

[OutThisLife]

Closing as implemented on main.

#80797 (merged, bdee48928) replaced the hand-written probe with _session_db_read_probe_statements(), derived from SCHEMA_SQL via schema_read_probe_statements() — so last_read_at and every future column are covered automatically. The _SESSION_DB_READ_PROBE_SQL constant this patches no longer exists on main.

Your diagnosis was right, and it's the fix that landed — just generalized so the probe can't go stale again the next time a column ships. Thanks for chasing it down.