Expose Codex subscription usage limits over the fold-codex transport

[PepijnSenders]

Downstream consumer: humanlayer/synclayer#2663

Why the change

HumanLayer needs to show Codex subscription usage-limit alerts the way it now does for Claude, but Fold exposed no usage-limit data, so this reads what the ChatGPT backend already returns and surfaces it in a provider-neutral shape.

Special things to note

• the data was already arriving and being dropped: the codex backend attaches usage state as response headers (x-codex-<primary|secondary>-used-percent / -window-minutes / -reset-at, plus x-codex-credits-*) on every reply, verified against a real 429 on this machine, so no new client, endpoint, or websocket was needed
• surfaced through a caller-supplied onUsageLimits callback, not a durable EventLog entry: the model descriptor produces a LanguageModel.Service with no EventLog access and the existing onStreamRetry seam is the only precedent; a durable entry would need cross-cutting runtime plumbing and is left as a follow-up
• reads the default codex limit family only (not per-model rateLimitsByLimitId) and fires on successful acquisition only; the 429 path keeps its existing retry-after handling

Change outline

A provider-neutral, best-effort snapshot type in fold-core, modeled on the existing Usage.

// packages/fold-core/src/EventLog/UsageLimits.ts
UsageLimits = {
  limitId?: string
  limitName?: string
  windows: Record<string, {           // keyed by a neutral id (primary/secondary)
    usedPercent?: number              // 0-100
    windowMinutes?: number            // window length; read this, not the key
    resetsAt?: number                 // epoch ms
  }>
  credits?: { hasCredits?: boolean; unlimited?: boolean; balance?: string }
}
decodeUsageLimits = Schema.decodeUnknownOption(UsageLimits)   // None when it does not conform

The codex client already returns the response at acquisition, so headers are read at that seam and handed to the callback.

 decorateCodexClient (packages/fold-codex/src/CodexModel.ts)
   acquire response over SSE (chatgpt backend /codex/responses)
+    usageLimitsFromCodexHeaders(response.headers)   // seconds -> ms, drop blank/out-of-range
+    onUsageLimits?(limits)                           // new CodexModelOptions callback, mirrors onStreamRetry

Files by responsibility.

 packages/fold-core/src/
+  EventLog/UsageLimits.ts        # provider-neutral UsageLimits schema + decodeUsageLimits
   index.ts                       # export it
 packages/fold-codex/src/
+  RateLimitHeaders.ts            # parse the codex header family into UsageLimits
   CodexModel.ts                  # read response.headers, invoke onUsageLimits
   index.ts                       # export the header parser

Tested: RateLimitHeaders.vi.test.ts (6 cases: primary+secondary parse, seconds-to-ms, partial windows, credits + limit name, no-headers None, unparsable/out-of-range dropped) and UsageLimits.vi.test.ts (4 cases: sparse windows, empty windows, percent bound, non-positive duration). Typecheck, oxlint, and oxfmt clean; the two failing CodexLive tests are pre-existing live-network cases that 429 because this account's limit is reached.

---

[CLAassistant]

All committers have signed the CLA.