diff --git a/.github/scripts/flake_report.py b/.github/scripts/flake_report.py
index 71398df825..67a6cc8ade 100755
--- a/.github/scripts/flake_report.py
+++ b/.github/scripts/flake_report.py
@@ -111,10 +111,12 @@ def cmd_count(args):
 _NON_TEST_TASK_FAILURE_RE = re.compile(r"Execution failed for task '([^']+)'")
 
 # The JVM's own crash banner -- reliable on every platform, including musl:
-# nothing but a real crash prints this.
+# nothing but a real crash prints this. The report file name only counts with
+# an actual pid: Gradle --info also prints every test JVM's command line, whose
+# -XX:ErrorFile=...hs_err_pid%p.log template is configuration, not a crash.
 _CRASH_RE = re.compile(
     r"A fatal error has been detected by the Java Runtime Environment"
-    r"|hs_err_pid"
+    r"|hs_err_pid\d+"
 )
 
 # Gradle reporting that a forked test JVM died. Reliable for the glibc/macOS
diff --git a/.github/scripts/tests/test_quarantine.sh b/.github/scripts/tests/test_quarantine.sh
index c7395c3d60..078e5bc238 100755
--- a/.github/scripts/tests/test_quarantine.sh
+++ b/.github/scripts/tests/test_quarantine.sh
@@ -725,6 +725,30 @@ assert d['final_attempt_cut_short'], d
 " "$CASE/out.json" || fail "a genuine musl crash was excused by the quarantine list"
 pass "a genuine crash banner still gates a quarantined musl failure"
 
+# Gradle --info prints each test JVM's full command line, which carries the
+# -XX:ErrorFile=...hs_err_pid%p.log template. That is configuration, not a
+# crash, and must not stop the list from excusing a quarantined failure.
+CASE="$TEMP_DIR/case-quarantined-errorfile-flag"
+mkdir -p "$CASE/flake-evidence/attempt-1"
+write_failure_xml "$CASE/flake-evidence/attempt-1" "com.dd.WobblyTest" "sometimesFails" "boom"
+cat > "$CASE/attempt.log" <<'EOS'
+Starting process 'Gradle Test Executor 1'. Command: /jdk/bin/java -XX:ErrorFile=build/hs_err_pid%p.log -Xmx1536m
+> Task :ddprof-test:testRelease FAILED
+EOS
+write_list "$CASE/list.txt" "$(entry com.dd.WobblyTest.sometimesFails PROF-1 "$(day_offset 30)")"
+python3 "$SCRIPTS/flake_report.py" --list "$CASE/list.txt" report \
+  --cell "el7-8-release-amd64" --evidence-dir "$CASE/flake-evidence" \
+  --final-attempt 1 --attempt-log "$CASE/attempt.log" \
+  --final-attempt-exit-code 1 --test-task-pattern test \
+  --out "$CASE/out.json" >/dev/null 2>&1
+python3 -c "
+import json,sys
+d = json.load(open(sys.argv[1]))
+assert d['gates'] is False, 'the ErrorFile flag must not read as a crash: %r' % d['gate_reason']
+assert not d['final_attempt_cut_short'], d
+" "$CASE/out.json" || fail "the -XX:ErrorFile hs_err_pid%p template was read as a crash"
+pass "the -XX:ErrorFile hs_err_pid%p template does not gate a quarantined failure"
+
 # Same intent, without the log saying so: the final attempt reached fewer tests
 # than an earlier one managed, so it stopped early.
 CASE="$TEMP_DIR/case-quarantined-but-short-run"
diff --git a/ddprof-test/src/test/java/com/datadoghq/profiler/sanity/SanityCheckTest.java b/ddprof-test/src/test/java/com/datadoghq/profiler/sanity/SanityCheckTest.java
index 0d1b783e25..aa2e660332 100644
--- a/ddprof-test/src/test/java/com/datadoghq/profiler/sanity/SanityCheckTest.java
+++ b/ddprof-test/src/test/java/com/datadoghq/profiler/sanity/SanityCheckTest.java
@@ -38,6 +38,37 @@ private Path newJfrPath(String prefix) throws Exception {
         return Files.createTempFile(rootDir, prefix, ".jfr");
     }
 
+    /**
+     * An upper bound, in MB, on the memory sanity check's available-memory figure: the smaller of
+     * {@code MemTotal} ({@code OS::getRamSize()}) and this cgroup namespace's root memory limit.
+     * The native check takes the minimum over the process's whole cgroup ancestry, which includes
+     * that root, so its figure can only be smaller than this.
+     */
+    private static long availableMemoryUpperBoundMb() throws Exception {
+        long bytes = Long.MAX_VALUE;
+        for (String line : Files.readAllLines(Paths.get("/proc/meminfo"))) {
+            if (line.startsWith("MemTotal:")) {
+                bytes = Long.parseLong(line.replaceAll("[^0-9]", "")) * 1024;
+                break;
+            }
+        }
+        for (String limitFile : new String[] {
+                "/sys/fs/cgroup/memory.max", "/sys/fs/cgroup/memory/memory.limit_in_bytes"}) {
+            Path path = Paths.get(limitFile);
+            if (Files.isReadable(path)) {
+                String value = new String(Files.readAllBytes(path)).trim();
+                // cgroup v2 spells "unlimited" as "max"; v1 uses a near-Long.MAX_VALUE number.
+                if (value.matches("[0-9]+")) {
+                    bytes = Math.min(bytes, Long.parseLong(value));
+                }
+            }
+        }
+        if (bytes == Long.MAX_VALUE) {
+            throw new IllegalStateException("MemTotal not found in /proc/meminfo");
+        }
+        return bytes / (1024 * 1024);
+    }
+
     /**
      * nosanity=true bypasses sanity checks. The profiler must start successfully on any host.
      */
@@ -98,7 +129,7 @@ void sanity_checks_run_once() throws Exception {
     }
 
     /**
-     * A forced -Xmx far larger than any real host's RAM makes the memory sanity check fail
+     * A forced -Xmx larger than the memory available to the process makes the memory sanity check fail
      * deterministically, regardless of the actual host resources. The check is advisory, so
      * the profiler must still start, and the JFR recording's settings must show the failure.
      */
@@ -120,7 +151,18 @@ void mem_sanity_check_failure_is_recorded_in_jfr() throws Exception {
             // to a fraction of -Xmx regardless of -Xms, so -Xms8m alone still eagerly
             // commits ~28g and OOMs before the sanity check runs. G1 sizes its initial
             // commit in fixed-size regions independent of -Xmx, avoiding that.
-            LaunchResult result = launch("profiler", Arrays.asList("-XX:+UseG1GC", "-Xmx900g", "-Xms8m"),
+            //
+            // JDK 8's G1 still allocates and clears card-granularity bitmaps spanning the
+            // whole reserved heap, one per parallel GC thread: measured on 8u504, a
+            // -Xmx900g fork touches ~225MB per ParallelGCThreads on top of a ~360MB base
+            // (1.3GB at 4 threads, 4GB at 16), enough to OOM-kill a 6GB CI container.
+            // -Xmx is therefore sized just past the check's available-memory figure (a
+            // 6GB-limited pod on a 380GB node needs ~7g, not 900g) -- its estimate is at least
+            // 1.3x -Xmx -- and a single GC thread keeps the per-thread cost bounded where
+            // that figure is large.
+            long xmxMb = availableMemoryUpperBoundMb() + 1024;
+            LaunchResult result = launch("profiler",
+                    Arrays.asList("-XX:+UseG1GC", "-XX:ParallelGCThreads=1", "-Xmx" + xmxMb + "m", "-Xms8m"),
                     "start,jfr,file=" + forkedJfr.toAbsolutePath(),
                     line -> LineConsumerResult.CONTINUE, line -> LineConsumerResult.CONTINUE);
             assertTrue(result.inTime, "forked JVM did not exit in time");