feat(v0.6-2): dispatcher state persistence -- cold-start, cooldowns, dedup LRU to SQLite

Closes Rule-20 dispatcher gap from audit doc v0.6-phase1-audit.md finding #1.
Pre-this-commit the cold-start anchor, 4 drop counters, per-toggle cooldown
map, and dedup OrderedDict all lived in Dispatcher instance memory and were
lost on every container restart.

v5.sql adds three tables:
  - dispatcher_state (singleton id=1): cold_start_anchor + 4 drop counters
  - dispatcher_cooldowns ((toggle,category,region) keyed): last_fired_at
  - dispatcher_dedup ((source,event_id) keyed): seen_at

Dispatcher refactor:
  - __init__ calls _restore_from_db -- counters, cold-start anchor, cooldown
    map, and dedup LRU (most-recent 10k by seen_at) all rehydrated from the
    three new tables
  - write-through on every mutation: _persist_state for counter/anchor,
    _persist_cooldown for cooldown UPSERT + 2*cooldown_s prune,
    _persist_dedup for dedup INSERT OR REPLACE + 7-day cleanup
  - in-memory caches stay authoritative on the fast read path
  - cumulative-since-install counters (NOT since-boot); LLM will be able
    to answer "we have dropped 47 stale events this week" after commit #5
    (env_reporter) lands
  - graceful degrade: missing v5 tables / persistence outage falls back to
    fresh in-memory state without crashing the constructor

Tests:
  - tests/test_dispatcher_persistence.py (17 tests): state restore on init,
    counter+cooldown+dedup survival across simulated restart, cooldown rearm
    within 2x window, dedup LRU rebuild caps at 10k, 7-day cleanup on insert,
    INSERT OR REPLACE on duplicate source+event_id, v5 migration idempotent,
    synthetic storm (50 events) -> restart -> replay (5 incl 1 duplicate)
    with the duplicate dedup-rejected and counters NOT resetting
  - tests/conftest.py (new): autouse MESHAI_DB_PATH redirection to per-test
    tmp file, so the dispatcher_*  tables on production /data dont get
    polluted by tests that construct Dispatcher() without an explicit fixture
  - tests/test_notification_toggles.py: _dispatch helper wipes dedup/cooldown/
    state tables between calls (per-call independence preserved; pre-v0.6-2
    in-memory-only Dispatcher reset naturally per instance)

Test count: 680 -> 697 (+17 new, 0 regressions).

Refs audit doc v0.6-phase1-audit.md finding #1.

tests/conftest.py

@@ -0,0 +1,31 @@
+"""Pytest fixture isolation for meshai persistence (v0.6-2).
+
+Before v0.6-2 the dispatcher held all state in instance memory, so tests
+that constructed `Dispatcher(...)` were inert w.r.t. SQLite. v0.6-2 made
+`Dispatcher.__init__` read/restore from the persistence layer, which by
+default points at `/data/meshai.sqlite`. Without isolation every test
+would now read+write production state, polluting across tests and across
+pytest invocations.
+
+This autouse fixture redirects `MESHAI_DB_PATH` to a per-test tmp file
+and clears the persistence-layer threading.local caches around each test.
+Existing tests that don't reference any fixture get isolation for free;
+tests that explicitly use a `db_path` (or similar) fixture can still
+override the env var inside their own fixture body -- last setenv wins.
+"""
+import pytest
+
+from meshai.persistence import close_thread_connection
+from meshai.persistence import db as _persistence_db
+
+
+@pytest.fixture(autouse=True)
+def _isolate_meshai_db(tmp_path, monkeypatch):
+    """Point MESHAI_DB_PATH at a tmp file per test."""
+    p = str(tmp_path / "meshai-test-isolated.sqlite")
+    monkeypatch.setenv("MESHAI_DB_PATH", p)
+    _persistence_db._initialised.clear()
+    close_thread_connection()
+    yield p
+    close_thread_connection()
+    _persistence_db._initialised.discard(p)