…race, fix flaky test (#328, #316)

Three skill auto-loading fixes:

1. Remove "netclaw" from GenericKeywords blacklist — the token was stripped
   from all skill keyword indexes, making identity queries like "What version
   of Netclaw" unable to trigger netclaw-manual. TF-IDF weighting already
   handles common tokens.

2. Apply fallback keywords immediately on startup before LLM enrichment
   completes. Previously, fallback keywords were skipped when a chat client
   provider was available, leaving a race window with zero keywords.

3. Purge stale keyword cache files during RescanAndUpdateIndex so old version
   caches don't cause misses that force re-enrichment.

Also fixes:

4. Flaky test Reconcile_disables_zombie_oneshot_reminders — PreStart's
   Self.Tell(ReconcileReminders) could arrive after the test's Ask, causing
   the startup reconcile to race with the zombie write. Fixed by draining
   the mailbox with two reconcile Asks before writing test data.

5. HandleReconcileAsync catch block now replies with zero-result ack so
   external Ask callers don't hang until timeout on transient failures.

…SW003)