feat: usage tracking — calls per 5h window, overflow to secondary provider

[justrach]

Parent: #257 — Layer 4 (Optimization)

What

Two responsibilities, split across layers:

Counting (in dispatch, Layer 0)

Every dispatch() call increments the counter for the backend used:

const UsageTracker = struct {
    claude_calls_5h: u32 = 0,
    codex_calls_5h: u32 = 0,
    window_start_ns: i128,
    pub fn record(self: *UsageTracker, backend: Backend) void { ... }
};

Load balancing (in resolve, Layer 1)

resolve() checks usage before picking a provider. If the resolved provider is near its quota limit (warn_at_percent from config), resolve() switches to the other provider:

// Inside resolve():
if (resolved.backend == .claude and usage.claudeNearLimit(providers.config)) {
    resolved.backend = .codex;
    resolved.model = mapToCodexEquivalent(resolved.model);
}

This only activates when both providers are available. Single-provider setups just warn.

Depends on

• resolve() function
• feat: provider dispatch — backend registry (claude/codex/sdk/local) #262 (dispatch — counting happens here)
• feat: provider detection — ProviderSet.detect() probes environment #264 (provider detection)

[justrach]

Blocked by: #262.

[justrach]

Blocked by: #264.

[justrach]

Blocked by: #265.