MeshCore reconnect persistence: implement 0=unlimited max-reconnect-attempts sentinel

connect() now translates a configured meshcore_max_reconnect_attempts of 0
(or <=0) into an effectively-unbounded count before handing it to the
meshcore library's ConnectionManager, so its retry loop never
permanently exhausts. config.py's comment already documented \"0 =
unlimited\" but that sentinel was never actually implemented -- literal 0
meant zero attempts, and the shipped default of 5 (at ~1s/attempt) gave up
after ~5s with no external supervisor to retry again, leaving MeshCore dead
until a manual container restart. Also flips the repo default from 5 to 0
so fresh deploys get unlimited retries without extra config.

Proven via a 60s forced-outage auto-recovery test: the link recovers from
any-length vnode/radio outage instead of giving up after ~5s.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

work/meshai/config.py

@@ -51,7 +51,7 @@ class ConnectionConfig:
     meshcore_host: str = ""                     # pyMC companion frame server host
     meshcore_port: int = 5050                   # pyMC companion frame server port
     meshcore_auto_reconnect: bool = True        # enable meshcore lib auto-reconnect
-    meshcore_max_reconnect_attempts: int = 5    # max reconnect attempts (0 = unlimited)
+    meshcore_max_reconnect_attempts: int = 0    # max reconnect attempts (0 = unlimited)
     meshcore_advert_interval_seconds: int = 86400  # periodic self-advert interval, 24h (0 = disabled)
     # MeshCore connection type: tcp | serial | ble (default tcp for back-compat)
     meshcore_conn_type: str = "tcp"

work/meshai/transport/meshcore_transport.py

@@ -43,6 +43,29 @@ _TELEMETRY_MAX_FAILURES = 3
 # inside the 300s reaper window with margin to spare.
 _KEEPALIVE_INTERVAL_SECONDS = 120
 
+# --- Reconnect persistence ("0 = unlimited" sentinel) ----------------------
+# config.py documents meshcore_max_reconnect_attempts as "0 = unlimited", but
+# that sentinel was never implemented here — the value was passed straight
+# through to the meshcore lib's ConnectionManager, whose retry loop is
+# `while self._reconnect_attempts < self.max_reconnect_attempts`. Taken
+# literally, 0 means ZERO attempts (immediate give-up), the opposite of
+# "unlimited", and any small bounded value (the shipped default is 5, at the
+# lib's flat 1s-per-attempt cadence) exhausts after ~5 seconds and then the
+# link stays down PERMANENTLY — there is no external supervisor for MeshCore
+# (see main.py's watchdog guard: "MeshCoreTransport manages its own
+# reconnect via the meshcore lib's auto_reconnect parameter"), so nothing
+# ever notices and retries again after that. A radio/vnode bounce longer
+# than ~5s (e.g. the 2026-08-02 device-perm heal test) killed MeshCore for
+# good until a manual container restart.
+#
+# Fix: honor the documented sentinel for real. connect() below translates a
+# configured 0 into this effectively-unbounded count, so the lib's own
+# proven-safe retry loop (still local TCP only, still ~1 attempt/sec, still
+# WITHOUT re-sending the connect-time self-advert — see
+# _post_reconnect_setup_async) just keeps going until the vnode/radio comes
+# back, no matter how long the outage lasts.
+_MC_RECONNECT_ATTEMPTS_UNLIMITED = 2_147_483_647
+
 # Numeric Cayenne-LPP type id → decoded field name.  Ids not in this map are
 # passed through as ``lpp_<id>`` so nothing is silently dropped.
 _LPP_ID_TO_FIELD = {
@@ -1976,6 +1999,16 @@ class MeshCoreTransport(MeshTransport):
         ble_address = getattr(self.config, "meshcore_ble_address", "")
         auto_reconnect = getattr(self.config, "meshcore_auto_reconnect", True)
         max_attempts = getattr(self.config, "meshcore_max_reconnect_attempts", 5)
+        if max_attempts <= 0:
+            # Documented sentinel (config.py: "0 = unlimited") — see
+            # _MC_RECONNECT_ATTEMPTS_UNLIMITED's docstring for why this was
+            # never actually unlimited before and why translating it here is
+            # the fix.
+            logger.info(
+                "MeshCoreTransport: meshcore_max_reconnect_attempts=%s (unlimited) -> %d",
+                max_attempts, _MC_RECONNECT_ATTEMPTS_UNLIMITED,
+            )
+            max_attempts = _MC_RECONNECT_ATTEMPTS_UNLIMITED
 
         # Human-readable target for logging — from the same descriptor that
         # self_info() reports, so the log and the API never disagree.