backend-listen prod: explicit SERVER_BATCH_ONLY_MOBILE_PERCENT=0

backend/charts/backend-listen/prod_omi_backend_listen_values.yaml

@@ -130,6 +130,8 @@ env:
     value: "0"
   - name: BATCH_PIPELINE_PERCENT
     value: "0"
+  - name: SERVER_BATCH_ONLY_MOBILE_PERCENT
+    value: "0"
   - name: BATCH_PIPELINE_SERVING_ENABLED
     value: "false"
   - name: BATCH_PIPELINE_WORKER_ENABLED

backend/docs/feature-flag-registry.md

@@ -277,7 +277,7 @@ and an explicit empty literal renders as `''`.
 | `PLAYBACK_ON_DEMAND_ENABLED` | Build playback only on open/play; stable generation/checksum source IDs; existing task/run-lock coalescing | backend | env | closed | {value: 'false', category: rollout} | false (backend-listen (chart), pusher (chart)); {value: 'false', category: rollout} (cloud_run/backend, cloud_run/backend-integration, cloud_run/backend-sync, cloud_run/backend-sync-backfill, gke/backend-listen, gke/pusher) | false (backend-listen (chart), pusher (chart)); {value: 'false', category: rollout} (cloud_run/backend, cloud_run/backend-integration, cloud_run/backend-sync, cloud_run/backend-sync-backfill, gke/backend-listen, gke/pusher) | — | graduate | on | 2026-11-15 | dazheng |
 | `REPEAT_WORK_REVISION_ENABLED` | Reuse existing sync outbox with 60min ceiling and exact successful snapshot-based notes/L1 dedupe; content-free call/final joins | backend | env | closed | {value: 'false', category: rollout} | false (backend-listen (chart), pusher (chart)); {value: 'false', category: rollout} (cloud_run/backend, cloud_run/backend-integration, cloud_run/backend-sync, cloud_run/backend-sync-backfill, gke/backend-listen, gke/pusher) | false (backend-listen (chart), cloud_run/backend, cloud_run/backend-integration, cloud_run/backend-sync, cloud_run/backend-sync-backfill, pusher (chart)); {value: 'false', category: rollout} (gke/backend-listen, gke/pusher) | — | pending | undecided | 2026-11-04 | dazheng |
 | `REVIEW_SURFACE_MODE` | Review queue, reversible agent changes and entity pages; off by default, on enables authenticated routes | backend | env | closed | on | on (backend-listen (chart), cloud_run/backend, gke/backend-listen) | on (backend-listen (chart), cloud_run/backend, cloud_run/backend-sync, gke/backend-listen) | — | pending | undecided | 2026-11-07 | dazheng |
-| `SERVER_BATCH_ONLY_MOBILE_PERCENT` | Default 0; iOS/Android unsalted owner-hash cohort nested in existing batch admission, with the same custody and language gates | backend | env | closed | — | — | — | — | pending | on | 2026-10-31 | dazheng |
+| `SERVER_BATCH_ONLY_MOBILE_PERCENT` | Default 0; iOS/Android unsalted owner-hash cohort nested in existing batch admission, with the same custody and language gates | backend | env | closed | — | — | 0 (backend-listen (chart)) | — | pending | on | 2026-10-31 | dazheng |
 | `SONIOX_CAPTURE_AXIS_DIAGNOSTICS` | Measure raw Soniox token clocks against wire PCM and the managed send ledger; default off, no placement changes | backend | env | closed | declared (gke/backend-listen, gke/pusher); false (cloud_run/backend, cloud_run/backend-integration, cloud_run/backend-sync, cloud_run/backend-sync-backfill) | false (backend-listen (chart), cloud_run/backend, cloud_run/backend-integration, cloud_run/backend-sync, cloud_run/backend-sync-backfill, gke/backend-listen, gke/pusher, pusher (chart)) | false (cloud_run/backend, cloud_run/backend-integration, cloud_run/backend-sync, cloud_run/backend-sync-backfill); true (backend-listen (chart), gke/backend-listen, gke/pusher, pusher (chart)) | — | pending | undecided | 2026-11-04 | dazheng |
 | `SONIOX_ELAPSED_AXIS` | Measure Soniox elapsed timestamps before enabling speaker windows | backend | env | closed | — | — | — | — | kill | off | 2026-11-04 | dazheng |
 | `SONIOX_IDLE_COHORT_CLOSE_SECONDS` | Treatment-arm close seconds for the uid-hash idle cohort; unset/off keeps the global value; floored at 20 s | backend | env | open | — | — | — | — | pending | undecided | 2026-11-21 | backend |
@@ -942,7 +942,6 @@ their code default (`fail` tells you which way a missing value resolves).
 - `SCREEN_TASK_STOP` — Stop screen-task gate and flagged extraction admission; default false; clients poll every 30 seconds with a 55-second lease (fail: inverted)
 - `SELFHEAL_MODE` — Conversation self-heal sweeper mode: off/detect-only/nudge/heal (fail: closed)
 - `SERVER_BATCH_ONLY_ENABLED` — Default on; consented supported-language platform sessions reserve durable batch ownership and open no managed live STT (fail: open)
-- `SERVER_BATCH_ONLY_MOBILE_PERCENT` — Default 0; iOS/Android unsalted owner-hash cohort nested in existing batch admission, with the same custody and language gates (fail: closed)
 - `SERVER_BATCH_ONLY_PLATFORMS` — Default macos,windows; comma-separated desktop platforms admitted to server batch-only capture; linux supported explicitly (fail: closed)
 - `SONIOX_ELAPSED_AXIS` — Measure Soniox elapsed timestamps before enabling speaker windows (fail: closed)
 - `SONIOX_IDLE_CLOSE_SECONDS` — Default on at 45 s; close paid Soniox transports after continuous active-VAD silence or 60 s with no client frames; 0 or off is the kill switch (fail: open)

backend/scripts/verify_pusher_cohost_env_diff.py

@@ -220,6 +220,10 @@
             "MOBILE_STANDARD_BATCH_ENABLED",
             "MOBILE_STANDARD_BATCH_PERCENT",
             "MOBILE_STANDARD_BATCH_UIDS",
+            # Only listen reads the mobile server-batch-only cohort: server_batch.selected via
+            # request_eligible and via eligible <- create_in_progress_conversation, both called only
+            # from routers/listen. Pusher's finalizer uses owned/has_audio and never admits capture.
+            "SERVER_BATCH_ONLY_MOBILE_PERCENT",
             # Managed listen-only STT rollout; pusher is not a live audio receiver.
             "PARAKEET_BATCH_PRESSURE_MIN_REPLICAS",
             "PARAKEET_BATCH_PRESSURE_POOL_HOST",

backend/tests/unit/test_backend_listen_helm_defaults.py

@@ -97,6 +97,37 @@ def test_backend_listen_helm_template_uses_runtime_project_for_google_cloud_proj
     assert 'name: TRANSLATION_NEGATIVE_CACHE_TTL\n              value: "21600"' in rendered
 
 
+def test_prod_listen_carries_an_explicit_zero_mobile_batch_only_percent():
+    """Ruling 73b217: the mobile batch-only cohort is a reviewed listen value, explicit 0 until a ramp PR changes it.
+
+    The value must reach the rendered container env (not only the values file), so a listen roll applies it.
+    """
+    values = _load_values(ENV_IDENTITY_DEFAULTS["prod"]["values_file"])
+    assert _env_value(values, "SERVER_BATCH_ONLY_MOBILE_PERCENT") == "0"
+    names = [entry.get("name") for entry in values.get("env", []) or [] if isinstance(entry, dict)]
+    assert names.count("SERVER_BATCH_ONLY_MOBILE_PERCENT") == 1
+
+    helm = shutil.which("helm")
+    if helm is None:
+        pytest.skip("helm is not installed")
+    rendered = subprocess.run(
+        [
+            helm,
+            "template",
+            "backend-listen",
+            str(CHART_DIR),
+            "-f",
+            str(ENV_IDENTITY_DEFAULTS["prod"]["values_file"]),
+            "--set-string",
+            "image.tag=abc1234",
+        ],
+        check=True,
+        capture_output=True,
+        text=True,
+    ).stdout
+    assert 'name: SERVER_BATCH_ONLY_MOBILE_PERCENT\n              value: "0"' in rendered
+
+
 def test_backend_listen_helm_template_requires_image_tag():
     helm = shutil.which("helm")
     if helm is None:

config/deployment-setting-classification.json

@@ -240,6 +240,7 @@
       "STT_LIVE_DEMAND_MIN_DWELL_SECONDS",
       "MOBILE_STANDARD_BATCH_ENABLED",
       "MOBILE_STANDARD_BATCH_PERCENT",
+      "SERVER_BATCH_ONLY_MOBILE_PERCENT",
       "MOBILE_STANDARD_BATCH_UIDS",
       "OWNER_IDENTITY_RETRY_COALESCE_ENABLED",
       "SPEAKER_GROUPING_CROSS_SCOPE_THRESHOLD",