Add governed agent learning proposal flow

[mdheller]

Summary

Adds the first Memory Mesh governed-learning proposal contract for durable agent learnings discovered during guarded AgentPlane sessions.

What changed

• Added schemas/agent-learning-proposal.schema.json
  • review-only proposal artifact
  • captures source session/task/repo/agent refs
  • captures destination scope/path/namespace
  • captures proposed playbook/memory diff
  • requires evidence refs and policy decision refs
  • requires review metadata
  • keeps durable writeback disabled by default
• Added examples/agent-learning/proposal.example.json
• Added scripts/validate_agent_learning_proposal.py
  • validates schema and semantic invariants
  • enforces review-only mode
  • enforces pending human review
  • blocks raw sensitive payload storage by default
  • enforces disabled writeback
• Added scripts/create_agent_learning_proposal.py
  • creates review-only proposal artifacts from agent/session evidence
  • never writes durable memory
• Added scripts/validate_agent_learning_proposal_generator.py
  • smoke-tests generation and validates generated proposal output
• Added make validate-agent-learning-proposal
• Added focused workflow .github/workflows/agent-learning-proposal.yml
• Updated README with the governed agent learning proposal surface

Why

Agent-discovered durable facts should become reviewable repo/project memory proposals, not chat residue and not silent writeback. This PR gives Memory Mesh the review-only artifact boundary needed by the SourceOS Agent Reliability Control Plane.

Validation

Expected validation:

python -m pip install jsonschema
make validate-agent-learning-proposal

Linked work

Refs #22
Refs SocioProphet/agentplane#92
Refs SourceOS Agent Reliability Control Plane workstream.