feat: add industrial embodied AI governance example

[carloshvp]

Summary

Adds a runnable industrial embodied-AI example in which a material-movement agent sends robot-cell requests through the live cMCP Runtime.

The example demonstrates three distinct outcomes:

• cMCP authorizes a declared workflow and the mock controller completes the motion
• cMCP denies an undeclared workflow before the controller is invoked
• cMCP authorizes the request but the independent controller rejects motion after the physical state changes

It also includes:

• a signed Agent Manifest binding the prompt, policy and tool declarations
• explicit workflow-scoped Cedar permits with default deny
• a mock robot-cell MCP server with fresh, authenticated and single-use state tokens
• a runtime-issued TRACE Trust Record and signed audit bundle captured from a real run
• local artifact, signature and audit verification
• controller regression tests for stale, modified and replayed state, speed limits and approved zones

Why

Embodied agents need a clear boundary between software authorization and machine safety. A cMCP allow decision can prove that a governed request was authorized under a specific policy, but it must not be interpreted as controller acceptance, physical completion or functional-safety certification.

The example makes that boundary concrete while exposing an honest follow-up gap: the current cMCP audit bundle binds request hashes and policy decisions, but does not populate a response hash for the controller outcome.

Developer-preview dependencies

requirements.txt pins the cMCP and TRACE commits used for the end-to-end rehearsal, plus cedarpy, until the summit release stack is available from PyPI.

Validation

• clean virtual environment install from requirements.txt
• pip check
• python validate_artifacts.py
• python -m unittest discover -s tests -v (7 passed)
• ruff check industrial-embodied-ai
• JSON and YAML parsing
• two consecutive end-to-end runs through cMCP
• cmcp verify with pinned policy/catalog hashes and the signed audit bundle

The committed TRACE fixture is intentionally software-only; it validates signatures, hashes and audit integrity without claiming hardware provenance. The agent supports --require-hardware for rehearsal on a supported TEE host.

[imran-siddique]

Reviewed in depth: security pass over every file (clean -- local-only network surface, no dynamic execution, dependency pins verified against the real agentrust-io repos), validate_artifacts.py and the 7 unit tests pass locally against the current verifier, the Cedar bundle parses and produces exactly the README's claimed decisions for declared/undeclared/missing workflows, and both catalog entries validate against the cmcp schema. The signed TRACE record and audit bundle check out as genuinely captured from a real run, and the noted response-hash gap in the audit bundle is accurate (filed for follow-up on our side).

I pushed one small commit to the branch: a .gitattributes pinning LF for the hash-bound artifacts -- on Windows checkouts with core.autocrlf=true the prompt was rewritten to CRLF and validate_artifacts.py failed on system_prompt_hash. Verified the failure and the fix locally.

One follow-up for later (no action needed now): requirements.txt pins cmcp-runtime to a feature-branch commit; once that branch merges we'll re-pin to a main commit so the SHA stays fetchable.

Excellent contribution -- the software-authorization vs. machine-safety boundary is exactly the right framing, and the example is honest about what a cMCP allow does and does not prove.