Harden rmap ordering and retry ticker advances

[raphael]

Summary

• order replicated-map writes by committed revision so destroy/reconnect recovery and SetAndWait acknowledgements stay consistent under concurrent updates
• route stream sink cleanup through the rmap destroy protocol and preserve destroy tombstones as intentional protocol state in tests
• retry distributed ticker advances after transient map write failures and add the repo AGENTS.md guidance in-tree

Test plan

• Run go test ./rmap ./streaming ./pool in a Redis-backed environment
• Verify rmap.Destroy() and reconnect keep revision ordering intact across replica resync
• Verify a transient ticker map write failure rearms the local timer and resumes ticking without unrelated map traffic

Notes

• I reviewed the existing local commit 3a2cc7f before opening this PR and did not find a blocker that required changes before shipping.
• Local go test ./pool currently fails during Redis test setup with EOF, so I did not claim a clean package test run in this environment.