fix(snmp): poll the MokerLink switch gently enough not to wedge it

stacks/observability/snmp-exporter/generator.yaml

@@ -83,9 +83,24 @@ modules:
   # NOTE: 32-bit counters only. Adding ifXTable (64-bit) is tracked in
   # docs/roadmap.md — on gigabit links ifInOctets wraps in ~34 seconds at
   # line rate, which the 60s scrape interval cannot detect.
+  #
+  # The tuning below is not cargo cult. This switch's management CPU has locked
+  # up under exporter polling before, taking its SNMP agent down while layer 2
+  # kept forwarding and the web UI kept answering — which presents as UDP/161
+  # timing out while ICMP and TCP/80 succeed. Recovery was a power cycle. What
+  # ran afterwards without wedging it used small GETBULK batches and a longer
+  # per-request timeout, so those values are pinned here rather than left to
+  # the exporter's defaults (max_repetitions 25, retries 3, timeout 5s).
+  #
+  # 5 * ~22 ifTable columns is a handful of round trips against a switch that
+  # cannot absorb 25. Worst case here is retries * timeout = 21s, comfortably
+  # inside the job's 45s scrape_timeout.
   mokerlink:
     walk:
       - 1.3.6.1.2.1.2.2  # IF-MIB::ifTable
+    max_repetitions: 5
+    retries: 2
+    timeout: 7s
 
   # HPE ProLiant DL360 Gen9 iLO (shiva, 10.0.30.10) — SNMPv2-MIB system group
   # plus the Compaq/HP Insight enterprise tree for hardware health.