chore: release v1.21.0-beta.4

[Soju06]

Summary

• Prepares beta release v1.21.0-beta.4 for the 1.21.0 stable train.
• Updates release-managed version files only; release-please remains the stable release owner.
• Merging this PR will publish v1.21.0-beta.4 only after the release-candidate validation checklist below is completed for this exact candidate SHA.
• Synced automatically from release-please PR chore(main): release 1.21.0 #1087; no manual beta workflow dispatch is required.

Changes since v1.21.0-beta.3

• fix(proxy): accept large websocket response.create and fail fast with 400 (fix(proxy): accept large websocket response.create and fail fast with 400 #1236) (02f1784)
• perf(request-logs): skip-scan unfiltered filter-option facets (perf(request-logs): skip-scan unfiltered filter-option facets #1239) (
  7c2cd9)
• perf(proxy): cut pre-upstream DB round trips on interactive turns (perf(proxy): cut pre-upstream DB round trips on interactive turns #1241) (
  b62797)
• perf(proxy): make SSE event framing linear instead of quadratic (perf(proxy): make SSE event framing linear instead of quadratic #1242) (
  ed9d4d)
• perf(clients): keep upstream connections and DNS across interactive turns (perf(clients): keep upstream connections and DNS across interactive turns #1245) (
  08e7a9)
• perf(dashboard): compress responses, cache assets, lazy-load charts (perf(dashboard): compress responses, cache assets, lazy-load charts #1243) (
  c01abc)
• perf(usage): aggregate the summary window in SQL, drop dead log refresh (perf(usage): aggregate the summary window in SQL, drop dead log refresh #1244) (
  22544b)
• perf(accounts): serve request-usage summaries from persistent rollup (perf(accounts): serve request-usage summaries from persistent rollup #1238) (
  f8d4be)
• feat(retention): opt-in pruning for request logs and usage history (feat(retention): opt-in pruning for request logs and usage history #1240) (
  637c7c)
• fix(db): serialize startup migrations across replicas (fix(db): serialize startup migrations across replicas and report schema-ahead correctly #1256) (
  0a7f35)
• fix(quota-planner): atomic warmup budget claims (fix(quota-planner): make warmup budget and dedup claims atomic across replicas #1260) (
  a8e12f)
• fix(balancer): persist rate-limit cooldowns so peer replicas honor them (fix(balancer): persist rate-limit cooldowns so peer replicas honor them #1261) (
  991be8)
• fix(deploy): repair broken multi-replica deployment artifacts (fix(deploy): repair broken multi-replica deployment artifacts #1263) (

• fix(proxy): fence bridge lease writes, evict fenced-out replicas, purge orphaned ring state (fix(proxy): fence bridge lease writes, evict fenced-out replicas, purge orphaned ring state #1259) (
  53f7b4)
• fix(caching): extend cross-replica cache invalidation bus to routing, selection, and settings caches (fix(caching): extend cross-replica cache invalidation bus to routing, selection, and settings caches #1255) (
  b7bf87)
• feat(replica-ops): document replica topology contract and land startup guardrails (feat(replica-ops): document replica topology contract and land startup guardrails #1264) (
  e9392f)
• feat(proxy): partition account concurrency caps across live replicas (feat(proxy): partition account concurrency caps across live replicas #1258) (
  d4f254)
• fix(reset-credits): make redemption replica-safe with durable ledger and claim (fix(reset-credits): make redemption replica-safe with durable ledger and claim #1257) (
  01188e)
• fix(proxy): queue bridged requests through response-create gate contention (fix(proxy): queue bridged requests through response-create gate contention #1266) (
  87fae4)
• feat(balancer): adapt selection and refresh to absent short-window quotas (feat(balancer): adapt selection and refresh to absent short-window quotas #1267) (
  9c1486)
• feat(planner): derive phase planning from observed short windows (feat(planner): derive phase planning from observed short windows #1268) (
  7fdbdc)
• perf(proxy): parse each SSE event once per layer, skip no-op reserialization (perf(proxy): parse each SSE event once per layer, skip no-op reserialization #1270) (
  f3cc5b)
• perf(proxy): detach stream-end persistence from the response path (perf(proxy): detach stream-end persistence from the response path #1269) (
  0b4363)
• perf(request-logs): cache the listing total per filter signature (perf(request-logs): cache the listing total per filter signature #1277) (
  27b982)
• perf(frontend): code-split dashboard routes (perf(frontend): code-split dashboard routes #1278) (
  3e4479)
• perf(frontend): self-host JetBrains Mono, drop Google Fonts (perf(frontend): self-host JetBrains Mono, drop Google Fonts #1279) (
  acfb29)
• feat(models): replicate leader-refreshed model registry via persisted snapshots (feat(models): replicate leader-refreshed model registry via persisted snapshots #1262) (
  fa3906)
• fix(automation): fall back to github.token when label sync hits rate limits (fix(automation): fall back to github.token when label sync hits rate limits #1272) (
  3ce5ba)
• feat(usage): ingest live rate-limit snapshots from proxied traffic (feat(usage): ingest live rate-limit snapshots from proxied traffic #1273) (
  75e7c2)
• feat(dashboard): hide expired primary windows instead of freezing them (feat(dashboard): hide expired primary windows instead of freezing them #1301) (
  165aa9)
• feat(ci): mint a dedicated GitHub App token for codex label sync (feat(ci): mint a dedicated GitHub App token for codex label sync #1302) (
  5470a5)
• fix(scheduling): harden scheduler leader election for multi-replica safety (fix(scheduling): harden scheduler leader election for multi-replica safety #1253) (
  13d3a3)

Release-candidate validation

Validated candidate: 3b3a93d

Complete these checks on the exact candidate SHA above before merging. The beta release guard and publish workflow fail while any item remains unchecked or the SHA is stale.

• Backend/unit/integration gates
• Frontend tests/build
• Wheel/package validation
• Docker/container smoke
• Live upstream/account smoke
• Live upstream/account smoke not required

Validation evidence (candidate 3b3a93d, validated 2026-07-15)

• Backend gates: full local suite on the exact candidate SHA — 5,218 passed, 9 skipped, 0 failed; ruff/format/ty/architecture-cap/openspec (43 specs) all clean. CI integration-bridge failure on attempt 1 (reconnects_after_clean_upstream_close, 409 vs 200) reproduced 0/13 locally and passed on rerun attempt 2 — flake, not a regression.
• Frontend: vitest 841/841 passed; production build clean on the candidate.
• Wheel: uv build → codex_lb-1.21.0b4 wheel+sdist; fresh-venv install reports __version__ 1.21.0-beta.4, server boots against sqlite, /health + /health/ready (db ok) + dashboard 200.
• Docker smoke: image built from candidate; booted against a migrated copy of an existing sqlite volume (Alembic upgraded to 20260713_020000_add_model_registry_snapshot), /health/ready ok. (The copied scratch volume was pre-corrupted — stamped past 20260613_000000 without the codex_installation_id column — repaired manually; not a candidate defect since the migration graph correctly skips already-stamped revisions.)
• Live upstream/account smoke PERFORMED (real Codex CLI 0.144.1 + real ChatGPT pro account, scratch beta4 container on the candidate image): WebSocket transport (/backend-api/codex/responses accepted), capacity_weighted selection + sticky codex_session + previous_response continuity resolution, gpt-5.4-mini and gpt-5.6-luna round trips both succeeded (exact-match replies, 1,296 + 7,468 tokens), request logs recorded, weekly-only (10080-min) usage normalized to primary and ingested — the adaptive-window path exercised live.

Install after publish

uvx --from "codex-lb==1.21.0b4" codex-lb
docker pull ghcr.io/Soju06/codex-lb:1.21.0-beta.4
helm install codex-lb oci://ghcr.io/soju06/charts/codex-lb --version 1.21.0-beta.4 --devel

Promotion

Merge the normal release-please PR to promote this beta-tested train to 1.21.0.