Jevgrep 0.4.4

TypeSafe requests now use token-aware pacing while keeping the default 32 concurrent requests. Each search reserves estimated input tokens before dispatch, reconciles reported usage, and respects the provider's token and request windows. Retries share those budgets; queue waits remain cancellable and do not consume the network timeout.

Native TypeSafe requests also allow longer response times to avoid premature retries. Retrieval, prompts, and the public skill are unchanged in this release.

Validated with the complete test suite, 40 installed-package journeys, and a live native TypeSafe retrieval with complete usage and no failed responses. This release does not claim a new benchmark cost reduction.