feat: support nested peer replies and retry failed runtime starts

[arkash20]

An agent handling a leased request can now read responses to its own helper request through peer recent(reply_to=...), without acknowledging unfinished work or acquiring another lease. If the runtime executable cannot start, the waker restores its previous marker so installation/permission repair can be retried; ambiguous failures after process creation still suppress duplicate prompts.

Stacked follow-up to #1666; base is feat/agent-collaboration. The paired Enterprise PR #1885 enforces the new messaging_scope policy and implements the response query. Its default is same-fleet messaging, with explicit admin opt-in for tenant-wide sends. Keep PostgreSQL delivery, one active lease, private stdio tokens, human oversight and explicit memory writes from the foundation PRs.

Validation:

• Missing-executable regression reproduced before the fix.
• 80 collaboration client tests passed; 70.29% aggregate package coverage.
• 26 native optional-entrypoint/auth boundary/remote-MCP tests passed.
• Ruff/format and all four wheel/sdist builds passed.
• Paired isolated stack passed dashboard (7), messaging/recovery (16), oversight (10), stdio pull/restart (5), and new two-client nested round-trip (3) checks.

Reading a response leaves its delivery queued for later normal acknowledgement. Same-fleet policy enforcement requires the paired Enterprise change. This is a preview follow-up, not universal runtime/GA qualification. A model-driven probe remains unrun because automatic approval review rejected external model execution with generated test credential configuration; deterministic runtime and real stdio tests did run.

Normal native CI is enabled for the stacked feat/agent-collaboration base in both repositories; it otherwise only ran on main/dev.

Final GitHub checks all passed on 7a314c829921042c1ac44e1e485ed977e5dd0b61: native CI run. Confirmed 2026-09-23. No review comments are outstanding.