diff --git a/docs/push-batch-investigation.md b/docs/push-batch-investigation.md
new file mode 100644
index 0000000..aea44fa
--- /dev/null
+++ b/docs/push-batch-investigation.md
@@ -0,0 +1,12 @@
+# Push batch lifecycle investigation
+
+Original failure: main commit 6e24bcbead3d352f305fb92bee9d894de441f554, run 34237080747, Debian full-suite context.
+The partial-transfer case entered lifecycle failure before the following batch failed accounting.
+
+Current probes require exact ordinary failure status and preserve lifecycle barriers. Eight fixed rounds stop on the first failure; no retry-to-green.
+Selective push-only runs have not reproduced the original failure. This document is an unknown dependency under the existing tests/run.sh mapping, intentionally exercising the existing full-suite fallback for this diagnostic PR. No workflow or selection rules are changed.
+
+Capture only Bash version, return codes and lifecycle maps. No credentials, full environment dump or global xtrace.
+A green full run is not proof of a fix. Retain original logs and distinguish fixture instrumentation from production behavior.
+
+Next evidence needed: first failing boundary and pre-teardown state under the full suite. Production code remains unchanged; PR 47 stays blocked pending an explained resolution.
diff --git a/tests/test-push.sh b/tests/test-push.sh
index b433c16..fbc675b 100644
--- a/tests/test-push.sh
+++ b/tests/test-push.sh
@@ -1386,13 +1386,38 @@ EOF
     runtime_has_published_worker_state && fail "sliding batch left worker-session state"
     runtime_has_worker_registration_state && fail "sliding batch left registration state"
     : > "$SUCCESS_FILE"; : > "$FAILED_FILE"; : > "$root/state/current"; : > "$root/state/max"
+    diagnose_batch_contract() {
+        printf "DIAG: bash=%s status=%s lifecycle_failed=%s worker_error=%s\n" "$BASH_VERSION" "$1" "$BATCH_WORKER_FAILED" "$BATCH_WORKER_ERROR" >&2
+        declare -p ACTIVE_WORKERS ACTIVE_WORKER_STATE_FILES ACTIVE_WORKER_STATE_STARTS REGISTERING_WORKERS WORKER_REGISTRATION_READY_FILES WORKER_REGISTRATION_RELEASE_FILES WORKER_REGISTRATION_STARTS WORKER_REGISTRATION_STAGE_FILES >&2
+    }
+    # Keep production call stacks intact; diagnose only after return.
+    assert_transfer_failure_contract() {
+        local label=$1 status=$2
+        if [[ $status != 1 || $BATCH_WORKER_FAILED != false || $BATCH_WORKER_ERROR != false ]]; then
+            diagnose_batch_contract "$status"
+            fail "$label must be ordinary transfer failure, not lifecycle/accounting failure"
+        fi
+        pass "$label preserves ordinary failure status"
+    }
+    # Fixed probes: stop on first failure, never reset lifecycle barriers.
+    batch_probe_deadline=$((SECONDS + 90))
+    for batch_probe in 1 2 3 4 5 6 7 8; do
+    (( SECONDS < batch_probe_deadline )) || fail "batch probe budget exhausted"
+    printf "DIAG: consecutive failure batch round=%s/8\n" "$batch_probe"
+    : > "$SUCCESS_FILE"; : > "$FAILED_FILE"; : > "$root/state/current"; : > "$root/state/max"
     SERVERS=(good1 bad1 good2)
-    assert_fail "partial server failure returns nonzero" run_transfer source destination
+    partial_status=0
+    run_transfer source destination || partial_status=$?
+    assert_transfer_failure_contract "partial failure" "$partial_status"
+    assert_eq 2 "$(wc -l < "$SUCCESS_FILE")" "partial failure completes both unrelated servers"
     assert_eq 1 "$(wc -l < "$FAILED_FILE")" "partial failure records one failed server"
     : > "$SUCCESS_FILE"; : > "$FAILED_FILE"; : > "$root/state/current"; : > "$root/state/max"
     SERVERS=(bad1 bad2)
-    assert_fail "all server failures return nonzero" run_transfer source destination
+    all_status=0
+    run_transfer source destination || all_status=$?
+    assert_transfer_failure_contract "all failures" "$all_status"
     assert_eq 2 "$(wc -l < "$FAILED_FILE")" "all failures record every server"
+    done
     cleanup_runtime
 )