chore(deploy): mirror the 2026-10-24 backend-sync costcut into deploy config

.github/actions/deploy-backend-stack/action.yml

@@ -5,7 +5,7 @@ description: >-
 
 inputs:
   backend_sync_memory:
-    description: 'backend-sync memory: 4Gi prod right-size (2026-10-24), 6Gi/8Gi C023 trial values; dev callers keep 8Gi'
+    description: 'backend-sync memory: 6Gi prod right-size (2026-10-24 costcut), 4Gi/8Gi fallback trial values; dev callers keep 8Gi'
     required: false
     default: '8Gi'
   firebase_probe_signer_service_account:
@@ -690,13 +690,16 @@ runs:
         project_id: ${{ inputs.project_id }}
         image: ${{ steps.service-images.outputs.backend }}
         no_traffic: true
+        # 2026-10-24 costcut mirror: prod backend keeps its live shape in source (it was
+        # revision-inherited only, so a fresh service would deploy defaults). Dev inherits.
         flags: >-
           --revision-suffix=${{ steps.image-tag.outputs.revision_suffix }}-base
           ${{ inputs.transcription_candidate_tag != '' && format('--tag={0}', inputs.transcription_candidate_tag) || '' }}
           ${{ inputs.deploy_profile == 'auto-dev' && format('--tag={0}', inputs.candidate_tag) || '' }}
           --remove-env-vars=MEMORY_ENABLED_USERS,MEMORY_CANONICAL_PROMOTION_CRON_ENABLED,MEMORY_CANONICAL_PROMOTION_CRON_INTERVAL_HOURS,MEMORY_CANONICAL_PROMOTION_FAST_TRACK_ENABLED,BASIC_TIER_MINUTES_LIMIT_PER_MONTH,BASIC_TIER_WORDS_TRANSCRIBED_LIMIT_PER_MONTH,BASIC_TIER_INSIGHTS_GAINED_LIMIT_PER_MONTH,BASIC_TIER_MEMORIES_CREATED_LIMIT_PER_MONTH,CONVERSATION_RELEVANCE_JEV_UID_ALLOWLIST,CONVERSATION_RELEVANCE_JEV_ENABLED,CONVERSATION_RELEVANCE_JEV_SHADOW_PERCENT,CONVERSATION_RELEVANCE_JEV_SHADOW_DAILY_CAP,CONVERSATION_RELEVANCE_KEEP_ALL_PERCENT,CONVERSATION_RELEVANCE_JEV_RESCUE,MEMORY_OWNER_JEV_FLIP_ENABLED,MEMORY_OWNER_JEV_FLIP_PERCENT,MEMORY_OWNER_JEV_SHADOW_PERCENT,MEMORY_OWNER_JEV_SHADOW_DAILY_CAP,CONVERSATION_SMART_MERGE_UID_ALLOWLIST,CONVERSATION_SMART_MERGE_WALLCLOCK_GAP_MODE,DESKTOP_PENDANT_CALL_LINK_MODE,DESKTOP_PENDANT_CALL_LINK_PERCENT,MEETING_NOTES_EPISODE_SELECTION,MEETING_NOTES_EPISODE_JEV_THRESHOLD,MEETING_NOTES_EPISODE_SELECTION_EFFORT,MEETING_NOTES_EPISODE_SELECTION_TIMEOUT_SECONDS,CONVERSATION_CALENDAR_CONTEXT_READ_ENABLED,CONVERSATION_OCR_CONTEXT_ENABLED,MEETING_NOTES_SCREEN_FRAMES_CONTEXT_ENABLED,SCREEN_FRAME_EGRESS_ENABLED,MEMORY_BELIEF_MODEL_ENABLED,FREE_TIER_MEMORY_SUPPRESSION,FREE_TIER_MEMORY_SUPPRESSION_COHORT,KNOWLEDGE_LEDGER_DRAIN_ENABLED,KNOWLEDGE_LEDGER_DRAIN_UID_ALLOWLIST,MEMORY_DAILY_MEMORY_SWEEP_ENABLED,MEMORY_DAILY_MEMORY_SWEEP_KILL_SWITCH,MEMORY_DAILY_MEMORY_SWEEP_MODEL_ENABLED,MEMORY_DAILY_MEMORY_SWEEP_MODEL_NAME,MEMORY_DAILY_MEMORY_SWEEP_MAX_MODEL_CANDIDATES,MEMORY_DAILY_MEMORY_SWEEP_MAX_MODEL_COST_USD,MEMORY_DAILY_MEMORY_SWEEP_COHORT_ENABLED,MEMORY_DAILY_MEMORY_SWEEP_COHORT_FLAG,MEMORY_DAILY_MEMORY_SWEEP_COHORT_NAME,MEMORY_DAILY_MEMORY_SWEEP_COHORT_TIMEOUT_SECONDS,MEMORY_DAILY_MEMORY_SWEEP_TIMEZONE_RECONCILIATION_ENABLED,MEMORY_DAILY_MEMORY_SWEEP_STAGGER_SECONDS,DEEPGRAM_SELF_HOSTED_ENABLED,DEEPGRAM_SELF_HOSTED_URL${{ steps.runtime-env.outputs.backend_remove_env_extra }}
           ${{ steps.runtime-env.outputs.cloud_run_flags }}
           ${{ steps.runtime-env.outputs.backend_identity_flags }}
+          ${{ inputs.runtime_env == 'prod' && '--cpu=2 --memory=8Gi --concurrency=20 --min-instances=3 --max-instances=100' || '' }}
         env_vars: |-
           ${{ steps.runtime-env.outputs.backend_env_vars }}
           OMI_BACKEND_RELEASE_SHA=${{ inputs.admitted_sha }}

.github/workflows/gcp_backend.yml

@@ -711,8 +711,8 @@ jobs:
           LADDER_V1_ARCHITECT_YEAR_PRICE_ID: ${{ vars.LADDER_V1_ARCHITECT_YEAR_PRICE_ID }}
         with:
           deploy_profile: manual
-          # 2026-10-24 right-sizing: prod backend-sync runs 4Gi (live 05:54Z); dev keeps 8Gi.
-          backend_sync_memory: ${{ vars.BACKEND_SYNC_MEMORY || (github.event.inputs.environment == 'prod' && '4Gi' || '8Gi') }}
+          # 2026-10-24 costcut (live 22:45Z): prod backend-sync runs 6Gi; dev keeps 8Gi.
+          backend_sync_memory: ${{ vars.BACKEND_SYNC_MEMORY || (github.event.inputs.environment == 'prod' && '6Gi' || '8Gi') }}
           # Read only for a development dispatch, so a production run never carries it.
           admitted_sha: ${{ needs.firestore_readiness.outputs.admitted_sha }}
           environment: ${{ github.event.inputs.environment }}

backend/deploy/ops/prod.yaml

@@ -67,8 +67,11 @@ cloud_run:
       HTTP_SYNC_JOBS_RUN_TIMEOUT: '1500'
       LANGCHAIN_TRACING_V2: 'false'
       MEMORY_MODE: 'off'
-    max_instances: 80
-    min_instances: 1
+    # 2026-10-24 costcut (live 22:48Z): cap 80 -> 50; warm floor 1 -> 5 so the sync
+    # pump keeps conversation-finalization headroom (min 5 x concurrency 20, the
+    # shape deploy-backend-stack sizes its queue against).
+    max_instances: 50
+    min_instances: 5
   backend-sync-backfill:
     env:
       SYNC_TASKS_QUEUE: sync-backfill-fresh
@@ -107,7 +110,9 @@ cloud_run:
       LANGCHAIN_TRACING_V2: 'false'
       MEMORY_MODE: 'off'
       TYPESENSE_HOST_PORT: '443'
-    min_instances: 1
+    # 2026-10-24: 3, matching the deploy flag (prod && 3); the old 1 fought every
+    # train and idled the projection pump between config-train applies.
+    min_instances: 3
 
 cloud_tasks:
   batch-quarantine-recovery:

backend/deploy/staging/deploy.sh

@@ -56,13 +56,13 @@ fi
 PY="${REPO_ROOT}/backend/.venv/bin/python"
 DEPLOY_LOG="${OMI_STAGING_DEPLOY_LOG:-$HOME/workspace/omi/data/run-2026-10-19-legacy-retire/lane02/deploys.log}"
 
-# Prod-shaped runtime per service (prod live Cloud Run, read 2026-10-19 on ro-prod; sync and
-# backfill memory follow the 2026-10-24 prod right-size to 4Gi). Replica counts are staging-sized
+# Prod-shaped runtime per service (prod live Cloud Run, read 2026-10-19 on ro-prod; sync memory
+# follows the 2026-10-24 costcut to 6Gi, backfill the right-size to 4Gi). Replica counts are staging-sized
 # by the overlay; these mirror concurrency, CPU, memory and timeout.
 shape_flags() {
   case "$1" in
     backend) echo "--concurrency=20 --cpu=2 --memory=8Gi --timeout=3600 --ingress=all" ;;
-    backend-sync) echo "--concurrency=20 --cpu=2 --memory=4Gi --timeout=3600 --ingress=all" ;;
+    backend-sync) echo "--concurrency=20 --cpu=2 --memory=6Gi --timeout=3600 --ingress=all" ;;
     backend-sync-backfill) echo "--concurrency=6 --cpu=2 --memory=4Gi --timeout=1500 --ingress=all" ;;
     backend-integration) echo "--concurrency=40 --cpu=2 --memory=2Gi --ingress=all" ;;
     *) return 1 ;;

backend/docs/runbooks/sync-ingest-archive.md

@@ -86,13 +86,14 @@ invoker SA and its endpoint-specific OIDC audience. Configure
 named task before the cohort ramp. Do not enable a worker without runtime-image,
 identity, index-readiness and route qualification.
 
-First deploy this SHA at the prod default memory (4Gi since the 2026-10-24
-right-sizing; development keeps 8Gi) through GitHub Actions. For the
-separate memory trial, the lead creates a tagged **no-traffic** 6Gi revision
-from that already-built same-SHA image, routes at most 5% to it and waits ≥12h.
-Only after qualification/full promotion set GitHub environment variable
-`BACKEND_SYNC_MEMORY=6Gi` for future deployments; the backend-stack action validates
-`4Gi|6Gi|8Gi`. Since the 2026-10-24 right-sizing, the manual workflow's fallback is `4Gi`
+First deploy this SHA at the prod default memory (6Gi since the 2026-10-24
+costcut promotion; development keeps 8Gi) through GitHub Actions. For a
+separate memory trial, the lead creates a tagged **no-traffic** revision at
+the trial size from that already-built same-SHA image, routes at most 5% to it
+and waits ≥12h. Only after qualification/full promotion change the prod
+fallback here (or set GitHub environment variable `BACKEND_SYNC_MEMORY`)
+for future deployments; the backend-stack action validates
+`4Gi|6Gi|8Gi`. Since the 2026-10-24 costcut, the manual workflow's fallback is `6Gi`
 for prod and `8Gi` for development, and the dev-only deploys (auto-dev) keep the action's
 `8Gi` default. The deployment remains through the existing release admission and
 GitHub Actions workflow. Rollback: redeploy the admitted image at the prior size.

backend/tests/unit/test_cloud_run_rightsizing_pins.py

@@ -0,0 +1,90 @@
+"""Contract pins for the 2026-10-24 prod Cloud Run right-sizing (the lead's costcut).
+
+The T4 deploy (image 646e54d7, run 40138604325) re-asserted backend-sync at
+4Gi / max 80: #23495 had mirrored the 05:54Z right-sizing, not the evening
+costcut. These pins hold every deploy source -- the workflow fallback, the
+deploy-action flags, the ops-train values and the staging shape -- to the
+costcut values, so the next train can no longer silently revert them.
+
+Live reference (ro-prod, 2026-10-24 22:50Z): backend 8Gi/cpu 2/concurrency 20/
+min 3/max 100; backend-sync 6Gi/cpu 2/min 5/max 50; backend-sync-backfill
+4Gi/cpu 2/concurrency 6/min 0 (cap redis_guard-bounded, live 16);
+backend-integration 2Gi/cpu 2/concurrency 40/min 3/max 50.
+"""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+from scripts.ops import drift_check as dc
+
+RIGHTSIZED_SERVICES = ('backend', 'backend-sync', 'backend-sync-backfill', 'backend-integration')
+REPO_ROOT = Path(__file__).resolve().parents[3]
+
+
+def test_workflow_backend_sync_memory_fallback_is_the_costcut_value():
+    # vars.BACKEND_SYNC_MEMORY is unset in both GitHub environments, so the fallback
+    # literal is what a prod deploy passes. Pinned as text because the drift model
+    # only sees the expression through _WORKFLOW_ENV_SCOPED: an edit that breaks the
+    # expression shape would otherwise pass silently with stale model values.
+    text = (REPO_ROOT / '.github' / 'workflows' / 'gcp_backend.yml').read_text(encoding='utf-8')
+    assert re.search(
+        r"backend_sync_memory: \$\{\{ vars\.BACKEND_SYNC_MEMORY \|\| "
+        r"\(github\.event\.inputs\.environment == 'prod' && '6Gi' \|\| '8Gi'\) \}\}",
+        text,
+    )
+
+
+def test_prod_deploy_flags_match_the_live_right_sizing():
+    model = dc.source_model('prod')
+
+    backend = model['backend'].scale
+    assert (backend['memory'], backend['cpu'], backend['concurrency']) == ('8Gi', '2', 20)
+    assert (backend['revision_min'], backend['revision_max']) == (3, 100)
+
+    sync = model['backend-sync'].scale
+    assert (sync['memory'], sync['cpu']) == ('6Gi', '2')
+    # max-instances is live-resolved by design (Resolve live sync maxScale); the cap
+    # itself is pinned in the ops file. The deploy never touches sync min.
+    assert sync['revision_min'] is None and sync['revision_max'] is None
+
+    backfill = model['backend-sync-backfill'].scale
+    assert (backfill['memory'], backfill['cpu'], backfill['concurrency'], backfill['revision_min']) == (
+        '4Gi',
+        '2',
+        6,
+        0,
+    )
+
+    integration = model['backend-integration'].scale
+    assert (integration['memory'], integration['cpu'], integration['concurrency']) == ('2Gi', '2', 40)
+    assert (integration['revision_min'], integration['revision_max']) == (3, 50)
+
+
+def test_ops_train_values_match_the_live_right_sizing():
+    model = dc.source_model('prod')
+
+    sync = model['backend-sync'].ops_scale
+    assert (sync['service_min'], sync['service_max']) == (5, 50)
+
+    # redis_guard owns the backfill cap between its bounds (live 16).
+    backfill = model['backend-sync-backfill'].ops_scale
+    assert backfill['service_min'] is None
+    assert backfill['service_max']['owner'] == 'redis_guard'
+    assert backfill['service_max']['floor'] <= 16 <= backfill['service_max']['ceiling']
+
+    # Equal to the deploy flag (prod && 3): the ops file must never fight the deploy.
+    integration = model['backend-integration'].ops_scale
+    assert (integration['service_min'], integration['service_max']) == (3, None)
+
+    backend = model['backend'].ops_scale
+    assert (backend['service_min'], backend['service_max']) == (None, None)
+
+
+def test_staging_shape_keeps_parity_with_prod():
+    prod = dc.source_model('prod')
+    staging = dc.source_model('staging')
+    for name in RIGHTSIZED_SERVICES:
+        findings = dc._scale_vs_prod(name, staging[name].scale, prod[name].scale)
+        assert not findings, f'{name}: {[f.line() for f in findings]}'

backend/tests/unit/test_drift_check.py

@@ -458,8 +458,8 @@ def test_prod_source_model_reads_deploy_flags_and_the_ops_file():
     assert integration.scale['revision_min'] == 3 and integration.scale['concurrency'] == 40
     assert integration.ops_env['SUMMARY_DEBT_REPAIR_ENABLED'] == 'true'
     # The workflow's env-scoped backend_sync_memory fallback reaches the source model, so a
-    # future regression of the prod 4Gi right-size is drift, not silence.
-    assert model['backend-sync'].scale['memory'] == '4Gi'
+    # future regression of the prod 6Gi costcut is drift, not silence.
+    assert model['backend-sync'].scale['memory'] == '6Gi'
     assert model['backend-sync-backfill'].scale['memory'] == '4Gi'
     assert model['backend-sync-backfill'].ops_scale['service_max'] == {
         'owner': 'redis_guard',

backend/tests/unit/test_render_backend_runtime_env.py

@@ -1,4 +1,4 @@
 """Renderer for backend Cloud Run runtime env."""
 
 import json
 import os
@@ -592,13 +592,16 @@
     # here only; do not copy onto backend.
     # backend-sync DOES carry --cpu=2 (#21668): its 6-8Gi memory requires >=2
     # CPU on Cloud Run, and 1-CPU deploys of 8Gi are rejected outright.
+    # 2026-10-24 costcut mirror: backend now pins its OWN live shape (prod-scoped
+    # 8Gi/2 CPU/20/3/100) instead of inheriting it revision-to-revision; --cpu=2
+    # is legitimate there for the same #21668 pairing, but every
+    # integration-specific flag below still must not leak onto it.
     integration_flags = _deploy_backend_stack_step_flags('deploy-backend-integration')
     backend_flags = _deploy_backend_stack_step_flags('deploy-backend')
     sync_flags = _deploy_backend_stack_step_flags('deploy-backend-sync')
     # 2026-10-24 right-sizing: prod integration stays warm; the dev project idles at 0.
     integration_min = "--min-instances=${{ inputs.runtime_env == 'prod' && '3' || '0' }}"
     for flag in (
-        '--cpu=2',
         '--memory=2Gi',
         '--concurrency=40',
         integration_min,
@@ -609,11 +612,16 @@
         assert flag not in backend_flags, flag
         # backend-sync owns its own scale-out cap (see test_sync_two_lane.py);
         # its value may legitimately equal integration's, so only the other
-        # integration-specific flags must stay off it. --cpu=2 is likewise
-        # legitimate on sync (#21668: 8Gi needs >=2 CPU).
-        if not (flag.startswith('--max-instances=') or flag == '--cpu=2'):
+        # integration-specific flags must stay off it.
+        if not flag.startswith('--max-instances='):
             assert flag not in sync_flags, flag
+    assert '--cpu=2' in integration_flags
     assert '--cpu=2' in sync_flags
+    # Backend's own prod shape (dev branch empty: dev keeps inheriting).
+    assert (
+        "${{ inputs.runtime_env == 'prod' && '--cpu=2 --memory=8Gi --concurrency=20"
+        " --min-instances=3 --max-instances=100' || '' }}"
+    ) in backend_flags
 
 
 VERTEX_PT_CONTRACT = 'Company-paid Gemini text and vision generation may use only the active model'