Stop gateway outages from paging java-chat as failed

src/main/java/com/williamcallahan/javachat/service/EmbeddingModelKeepAlive.java

@@ -27,7 +27,15 @@ public class EmbeddingModelKeepAlive implements HealthIndicator {
     /** Probe latency above this marks a slow provider response without inferring its remote cause. */
     private static final long SLOW_PROBE_THRESHOLD_MILLIS = 5_000L;
 
-    /** Escalates only after a probe condition repeats on the next observation. */
+    /**
+     * Escalates only after a probe condition repeats on the next observation.
+     *
+     * <p>The escalation stays at WARN even when repeated: a background probe failure is a
+     * monitoring signal, never a user-facing request failure, and ERROR lines feed the
+     * critical java-chat-error-log Grafana alert. Provider-health paging is owned by the
+     * dependencies health group (java-chat-dependencies-degraded) and the gateway's own
+     * availability rules, so a routine gateway redeploy must not page here.</p>
+     */
     private static final int REPEATED_PROBE_ALERT_COUNT = 2;
 
     private static final long NANOS_PER_MILLISECOND = 1_000_000L;
@@ -178,7 +186,7 @@ private void recordFailure(
                     .log(() -> "event=embedding_model_probe_failed outcome=failure model=" + logSafeModelName
                             + " durationMs=" + probeDurationMillis + " consecutiveFailures=" + consecutiveFailureCount);
         } else if (consecutiveFailureCount == REPEATED_PROBE_ALERT_COUNT) {
-            log.atError()
+            log.atWarn()
                     .setCause(embeddingUnavailableException)
                     .log(() -> "event=embedding_model_probe_failure_loop outcome=failure model=" + logSafeModelName
                             + " durationMs=" + probeDurationMillis + " consecutiveFailures=" + consecutiveFailureCount);

src/main/resources/application-dev.properties

@@ -39,9 +39,17 @@ spring.thymeleaf.cache=false
 logging.level.org.springframework.boot.devtools=DEBUG
 
 # Development-specific overrides
-# More verbose logging for development
+# More verbose logging for development.
+#
+# `dev` is also the profile the deployed dev.javachat.ai runs under, so anything set here ships
+# to Loki. org.springframework.web at DEBUG was 99.5% of this service's log volume — a 5,000-line
+# sample over 3h on 2026-08-04 held 1,991 DispatcherServlet + 1,990 HttpEntityMethodProcessor +
+# 994 RequestResponseBodyMethodProcessor lines against 24 application lines, essentially all of it
+# request plumbing for the 15s /actuator/prometheus scrape. It buried the signal that matters
+# (EmbeddingModelKeepAlive, which the java-chat Loki alert rules parse) and cost 73,705 lines/24h
+# against production's 1,233. Raise it locally when debugging Spring MVC itself, e.g.
+# `--logging.level.org.springframework.web=DEBUG`, rather than for every deployed request.
 logging.level.com.williamcallahan.javachat=DEBUG
-logging.level.org.springframework.web=DEBUG
 logging.level.org.springframework.ai=DEBUG
 
 # CORS for Vite dev server (port 5173) + local Spring Boot (port 8085) + deployed dev host.

src/test/java/com/williamcallahan/javachat/service/EmbeddingModelKeepAliveTest.java

@@ -124,7 +124,8 @@ void repeatedFailuresEscalateOnceAndRecoveryRestoresHealth() {
 
         assertEquals(Status.DOWN, keepAlive.health().getStatus());
         assertEquals(1, eventCount(Level.WARN, "event=embedding_model_probe_failed"));
-        assertEquals(1, eventCount(Level.ERROR, "event=embedding_model_probe_failure_loop"));
+        assertEquals(1, eventCount(Level.WARN, "event=embedding_model_probe_failure_loop"));
+        assertEquals(0, eventCount(Level.ERROR, "event=embedding_model_probe_"));
 
         keepAlive.retryUnavailableEmbeddingModel();