[v2] Wire v2 Lattices into SafetyEngine Execution Pipeline

[SHAURYASANYAL3]

Phase 5: Integration & Migration

The Context

With all structural components (Capability Lattice, State Lattice, Cognitive Lattice) built, it is time to hook them into the core runtime of AgentWatch.

Tasks

• Open agentwatch/core/safety.py.
• Update the primary execution interceptor to route proposed actions through the new pipeline sequentially:
  1. CapabilityLattice.evaluate(action)
  2. StateLattice.simulate_and_check(action)
  3. CognitiveLattice.evaluate_health(session)
• Ensure that a failure at ANY level returns an immediate LatticeVerdict.BLOCKED or equivalent, with strict ~0.1ms latency budget.

Part of Epic #651

[prasiddhi-105]

Hiii please assign this issue to me under SSOC'26

[SHAURYASANYAL3]

@prasiddhi-105 Thanks for volunteering! Same as the other issue—since you're a newer contributor, could you drop a quick outline of how you plan to tackle this before we assign it?

[prasiddhi-105]

Thanks! @SHAURYASANYAL3 Here's my planned approach:

• I'll first review agentwatch/core/safety.py to understand the current execution interceptor and identify the appropriate integration point.
• I'll integrate the new lattice pipeline in the required order:
  1. CapabilityLattice.evaluate(action)
  2. StateLattice.simulate_and_check(action)
  3. CognitiveLattice.evaluate_health(session)
• I'll implement short-circuit evaluation so that if any stage returns a blocked verdict, the pipeline immediately returns LatticeVerdict.BLOCKED (or the project's equivalent) without executing the remaining stages.
• I'll keep the changes minimal and preserve the existing execution flow while ensuring the implementation remains lightweight to align with the ~0.1 ms latency target.
• Finally, I'll verify both the success and failure paths to ensure the pipeline behaves as expected.

Please let me know if you'd like me to follow any additional architectural conventions.

[SHAURYASANYAL3]

Your approach looks solid. You're officially assigned! Since our contributor rules limit everyone to 2 active issues at a time, you are currently at your cap. Let us know if you need any pointers while implementing.