v0.5.2

Bug-fix release: on self-hosted PG deployments the provider-limits
(balance/quota) cache froze, sync 500'd, and the dashboard read an empty
map even though rows were present in Postgres. Also fixes a 503 admission
regression from tight heavy-request limits.

Fixed

• Cloud-detection short-circuit: the probe (typeof globalThis.caches === "object") fired on every standalone server because Next 16 polyfills the
  Web Cache API, making balance/limit cache reads return empty and writes
  silently no-op. Cloud semantics now require the edge runtime
  (NEXT_RUNTIME=nodejs excluded) or explicit OMNIROUTE_CLOUD=true.
• SiliconFlow balance guard: the /v1/user/info payload now reports a
  degenerate all-string-zero shape as unavailable instead of a genuine 0
  balance.
• INSERT OR REPLACE upsert broke under webpack module duplication: the
  Postgres adapter (and its module-scoped pk-cache) is bundled into several
  chunks, each holding its own copy; a route served through a different chunk
  held a cold pk-cache and degraded to ON CONFLICT DO NOTHING, so
  already-cached connections never updated. Each adapter now warms its pk-cache
  lazily on first use and re-translates to a real ON CONFLICT (...) DO UPDATE.
• Named-placeholder regression: the adapter rework dropped the SQL rewrite
  mapping named SQLite placeholders (@name/:name) to positional $N, so
  queries like is_active = @isActive hit Postgres with column "isactive" does not exist and provider-limits sync returned 500. buildBinder now
  returns both the positional SQL and the binder.
• Read path surfaced an empty map: getSanitizedCachedProviderLimitsMap
  consumed await getAllProviderLimitsCache() synchronously (no await), so
  Object.keys(caches) ran against a Promise and the map degraded to {}
  even though rows were present in Postgres.
• Admission lease TTL: the chat heavyweight admission lease is now
  force-released after a configurable TTL
  (OMNIROUTE_CHAT_HEAVY_LEASE_TTL_MS, default 10 min) as a defensive
  backstop against a stuck upstream SSE stream permanently occupying one of
  the heavy in-flight slots, which would otherwise push legitimate concurrent
  heavy traffic into retryable 503s (chat_admission_busy).

Full Changelog: v0.5.1...v0.5.2