agent-message delivery: unbounded waits and no delivery idempotency (duplicate execution)

[Dmatut7]

Verified against main (fb2db8ee1).

1. Four waits on the agent-message path have no timeout. A send to a session that is mid-hydration or suspended parks the sender's whole turn indefinitely; on our line we bound them (15-120s, settings-driven) with retryable factual errors, after production cases of turns stuck forever.

2. No idempotency on delivery → retries execute twice. Cross-process send times out at 30s on the caller while the supervisor leg has a 24h budget: the timeout errors but the message still lands, so a model retry delivers the same instruction twice. We fixed it with a sender-minted message id: first errored send is recorded as uncertain, same-key resend fails closed with an actionable error (receipt via agent_observe, or mint a new id).

Both have red-first regression tests on our line; details in https://github.com/Dmatut7/prime-agent-rlm/blob/merge/repl-kernel/docs/fork/upstream-feedback-20260911.md (§2.4/§2.5).