Repair MCP simulator and align public security, hypotheses and roadmap

[niksacdev]

The public README described a multi-agent robotics workflow, but the repository contained only a synthetic telemetry prototype with broken MCP startup/client APIs. This change makes the local demo runnable and states explicitly that Phase 1 is partial and no research hypothesis is validated.

• Replace import-time package installation and unsupported networking APIs with an SDK stdio client/server lifecycle. Validate simulated updates atomically, refresh timestamps, reject invalid fields/non-finite values, and bound IDs and cached records. Remove fabricated health metrics and physical-calibration claims.
• Refresh locked dependencies on maintained MCP 1.x; add protocol/regression tests, CI with immutable action references, uv/Actions Dependabot updates, and locked devcontainer setup.
• Rewrite README and all four existing planning documents against implementation evidence. Add SECURITY.md, a dated privacy/security audit, and a sourced Physical AI market assessment. Recommend a focused operator-diagnosis evaluation before A2A, cloud deployment or model tuning.

Validation: 17 tests pass locally on Python 3.11, including a real MCP subprocess round trip and rollback invariants. Demo, compilation, local documentation links, lock consistency and whitespace checks pass. pip-audit reports no known vulnerabilities across 59 installed packages, versus 16 distinct package/advisory findings across six original locked packages (31 raw records include duplicates). All 902 locked artifact hashes were verified against official PyPI; committed URLs use official PyPI. CI also tests Python 3.10 and 3.12.

Audit limits and follow-up: no credential-pattern matches in 10 fetched historical commits / 41 blobs; a historical work email remains exposed (not reproduced here). GitHub secret scanning, push protection, Dependabot security updates and private vulnerability reporting were disabled at inspection; repository settings were not changed. Main has effective PR/force-push/deletion rules. Devcontainer build, real robots and cloud integrations were not tested. The sample remains synthetic, local and unauthenticated. See docs/security_review_2026-09.md and docs/market_assessment_2026-09.md for evidence and remaining boundaries.

Do not merge as part of this audit.