[P0] Publish the AgentMesh Reliability Model and chaos benchmark

[0YHR0]

Problem

AgentMesh has stronger failure semantics than typical Agent frameworks, but they remain mostly
architecture claims and unit/integration tests. Users need reproducible evidence that the system
converges after realistic failures.

Proposal

Publish AgentMesh Reliability Model v1 and a runnable agentmesh-chaos benchmark. Start with a
small deterministic workload and failure matrix, then report correctness and recovery metrics.

Required initial injections:

• duplicate Outbox/Redis delivery;
• worker death before and after business commit;
• lease expiry and fenced stale completion;
• Redis interruption and recovery;
• checkpoint/business-state disagreement;
• MCP write succeeds but response is lost;
• approval/reconciliation process restart.

Acceptance criteria

• The reliability model documents invariants, crash windows, authoritative state, and convergence.
• Benchmark runs from one command in CI or a dedicated Compose profile.
• Every scenario asserts lost tasks, duplicate irreversible effects, stuck nonterminal state, and
  invariant violations—not only process exit status.
• Results include recovery latency, throughput, outbox lag, duplicate suppression, and unresolved
  outcome count in machine-readable JSON plus a human report.
• A checked-in baseline is tied to commit, environment, and scenario versions.
• CI runs a bounded smoke matrix; a larger repeatable profile is documented.
• Existing rollout proposal [Proposal] Add first-class rollout groups for isolated, concurrent, verifiable agent execution #26 reuses the same isolation/evaluator/failure evidence where relevant.

Non-goals

• Marketing “production ready” from one laptop result.
• PostgreSQL HA certification in the first increment.
• Model-quality benchmarking or RL training.

[0YHR0]

PR #144 defines R1-R12 invariants, named crash windows, an independent oracle, required CI/extended chaos scenarios, hard qualification gates, and machine-readable evidence.