Make MCP execution crash-recoverable

[bordumb]

Outcome

A gateway can die after durable reservation or provider entry and a replacement process can recover the exact native-owned execution without re-entering the provider or reconstructing Auths semantics.

Changes

• add authenticated native checkpoints for reserved, possible-effect, and receipt-pending states
• require execution stores to commit checkpoints with reservation and provider entry
• add explicit matching auths.recover(...) APIs in TypeScript and Python
• persist one authority validity epoch with the recoverable development manifest so independent processes derive the same authority and execution identity
• make development stores reject old disposable execution state, validate reference ownership, close completed checkpoints, and accept only byte-identical receipt replay
• add process-kill recovery tests in both SDKs

Safety

• recovery records remain opaque and authenticated by Rust
• recovery reauthorizes the exact action and derives the stable execution identity before loading pending state
• provider-entry recovery always reconciles; it never blindly invokes again
• stale or completed references fail closed

Validation

• TypeScript build passes
• Rust and Python formatting applied
• authoritative validation delegated to GitHub CI per repository policy