Avoid exhaustion cooldown for member prose

fleetd/src/main/java/dev/ltms/fleet/inject/CompletionResolver.java

@@ -380,7 +380,9 @@ public final class CompletionResolver implements TurnListener {
                         + "matched the profile's exhausted pattern): {}", target, reason);
                 // CB-578 stage B: only on the resolution that actually won the race — a late
                 // duplicate must never quarantine a credential twice for one refusal.
-                exhaustionSink.onExhausted(target, reason);
+                if (startsWithExhaustion(matchedLine, exhausted)) {
+                    exhaustionSink.onExhausted(target, reason);
+                }
             }
             return;
         }
@@ -478,7 +480,9 @@ public final class CompletionResolver implements TurnListener {
                         + "usable assistant block; no fleet_reply): {}", target, reason);
                 // CB-578 stage B: only on the resolution that actually won the race — a late
                 // duplicate must never quarantine a credential twice for one refusal.
-                exhaustionSink.onExhausted(target, reason);
+                if (startsWithExhaustion(matchedLine, exhausted)) {
+                    exhaustionSink.onExhausted(target, reason);
+                }
             }
             return true;
         }
@@ -614,19 +618,38 @@ public final class CompletionResolver implements TurnListener {
     }
 
     /**
-     * True when the error pattern begins the matched pane line, rather than appearing in prose.
+     * True when the pattern appears before the first sentence ending in the matched pane line, rather
+     * than after a member has started prose about it.
      *
      * <p>Leading terminal chrome is skipped first — box-drawing characters, bullets, gutter bars and
-     * spaces. #339 introduced this check with a bare {@code lookingAt}, and that rejected a genuine
-     * error line rendered as {@code "| 503 Service Unavailable: ..."}: the send still failed, but the
-     * credential outage was never recorded. That is the false negative #339's own invariant 3 called
-     * worse than the false positive it set out to fix — measured with a throwaway probe on the
-     * raw-scrape path, which is exactly the path whose comment says to expect leading chrome.
+     * spaces. Exhaustion patterns often name only the decisive words in a provider message, such as
+     * {@code "usage limit has been reached"}; they do not include its leading {@code "The"}. A bare
+     * {@code lookingAt} would therefore reject that genuine refusal, including one behind terminal
+     * chrome.
      *
-     * <p>Skipping only a leading run of non-letter, non-digit characters keeps the fix's intent. A
-     * member's prose ({@code "I checked the retry path. An API Error: makes it back off."}) still
-     * does not match, because there the pattern sits after words, not after chrome.
+     * <p>A member's prose about a refusal normally follows a completed sentence. That sentence ending
+     * is enough to keep it from reaching a cooldown sink while the send still fails with the full pane
+     * tail. This is deliberately less strict than the backend-error check because exhausted patterns
+     * may start after a provider's leading words.
      */
+    private static boolean startsWithExhaustion(String line, Pattern pattern) {
+        int i = 0;
+        while (i < line.length() && !Character.isLetterOrDigit(line.charAt(i))) {
+            i++;
+        }
+        var matcher = pattern.matcher(line);
+        if (!matcher.find()) {
+            return false;
+        }
+        for (int prefix = i; prefix < matcher.start(); prefix++) {
+            if (".!?".indexOf(line.charAt(prefix)) >= 0) {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    /** True when an error pattern begins the matched pane line after optional terminal chrome. */
     private static boolean startsWithBackendError(String line, Pattern pattern) {
         int i = 0;
         while (i < line.length() && !Character.isLetterOrDigit(line.charAt(i))) {

fleetd/src/test/java/dev/ltms/fleet/inject/CompletionResolverTest.java

@@ -484,6 +484,27 @@ class CompletionResolverTest {
 
     // --- CB-578 stage A: backend-exhausted classification ---------------------------------
 
+    @Test
+    void aNormalMemberReportMentioningTheExhaustionPatternDoesNotNotifyTheSink() {
+        String block = "⏺ I reviewed capacity handling. The usage limit has been reached means no more work can start.\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertEquals(Rendezvous.Kind.BACKEND_EXHAUSTED, waiter.getNow(null).kind(),
+                "a matching report still fails the send as exhausted");
+        assertTrue(waiter.getNow(null).text().contains("I reviewed capacity handling."),
+                "the exhausted result keeps the whole matched pane line");
+        assertTrue(notified.isEmpty(),
+                "a normal report mentioning an exhaustion pattern must not quarantine a credential");
+    }
+
     @Test
     void classifiesAMatchingScrapeAsBackendExhaustedInsteadOfACompletedReply() {
         String block = "⏺ Working on it...\nThe usage limit has been reached. Try again later.\n❯ ";
@@ -534,6 +555,25 @@ class CompletionResolverTest {
                 "the sink is told the matched reason: " + notified.get(0));
     }
 
+    @Test
+    void aRealExhaustionBehindTerminalChromeStillNotifiesTheSink() {
+        String block = "⏺ │ The usage limit has been reached. Try again later.\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertEquals(Rendezvous.Kind.BACKEND_EXHAUSTED, waiter.getNow(null).kind(),
+                "a real exhaustion must still fail the send as exhausted");
+        assertEquals(1, notified.size(),
+                "a real exhaustion behind terminal chrome must reach the sink");
+    }
+
     @Test
     void aLosingBackendExhaustedClassificationNeverNotifiesTheExhaustionSink() {
         // The waiter was already resolved (e.g. by the worker's own reply) before this scrape landed —