ci(families): add isolation review guardrails

[chaofengw-nv]

Background

PR NVIDIA/TensorRT-Model-Connect#1093 makes the model family the unit of ownership and fault isolation. Static architecture checks catch direct source dependencies, but they cannot prove runtime/source independence or consistently identify semantic family coupling introduced through shared tooling, benchmarks, examples, and validation code.

Exit Criteria

• A selected family's GPU/TRT and E2E tests run without sibling-family source visibility.
• Its runtime proof exposes only core, the selected backend, and that family's DSO.
• Automated review preserves intentional model-specific duplication and evaluates family ownership, shared semantic neutrality, benchmark/validation integrity, and shared-change blast radius on every push.
• Semantic checks remain advisory while their signal is calibrated; this PR does not make AI review the merge authority.

Implementation

• Run each selected family proof from a temporary source root containing only core/builder, root test configuration, and the owning family. Reject family symlinks and disable user-site imports so sibling imports and file reuse fail at execution time.
• Retain the existing isolated runtime root containing only core, TensorRT backend, and libtrtmc_model_<family>.so.
• Add REVIEW.md as the repository-wide Standards/Spec review contract for automated and human reviewers.
• Add four warning-level CodeRabbit custom checks plus focused instructions for family, tooling, benchmark, example, and validation paths. Reviews rerun on every push and report PASS, BLOCK, or HUMAN REVIEW REQUIRED.
• Add source-quality regression tests for isolation behavior and for the required CodeRabbit policy.

Validation

• ruff check --config ruff.toml tools/ci/context.py tools/ci/process.py tools/ci/e2e.py tools/tests/test_new_ci.py: passed.
• PYTHONPATH=core/builder:apps/benchmark:. python3 -m pytest tools/tests/test_new_ci.py -q: 26 passed.
• CI_BASE_REF=github/pr/1093 PYTHONPATH=core/builder:apps/benchmark:. python3 -m tools.ci pipeline source-quality: 107 passed on head 960389b955e5c4b822efec1121e075e6bfd333f1.
• PYTHONPATH=core/builder:apps/benchmark:. python3 -m pytest -q -m "not gpu and not trt" -p no:cacheprovider: 339 passed, 274 skipped, 5 deselected on the same head.
• .coderabbit.yaml validated against the current CodeRabbit v2 JSON schema.
• Target-hardware E2E was not run locally; GitHub and protected CI remain required evidence.

Notes For Future Readers

• This is a stacked PR based on the exact current head of refactor: isolate model families end to end NVIDIA/TensorRT-Model-Connect#1093 and must merge into agent/ai-native-horizontal-scale before refactor: isolate model families end to end NVIDIA/TensorRT-Model-Connect#1093, or be rebased onto main after refactor: isolate model families end to end NVIDIA/TensorRT-Model-Connect#1093 merges.
• Review REVIEW.md and .coderabbit.yaml first, then tools/ci/e2e.py, followed by the regression tests.
• CodeRabbit checks intentionally start in warning mode with request_changes_workflow: false. Promote a check to error only after real-PR calibration demonstrates acceptable signal.