chore: add input validation

[AlexanderBarabanov]

Summary

• src/physicalai/inference/component_factory.py
  • block path traversal in artifact paths via added _safe_resolve() to reject ../../ style values read from manifests
  • cap recursive component instantiation at _MAX_COMPONENT_DEPTH = 10 to prevent stack exhaustion from nested manifest specs (this issue has been confirmed during fuzzing session)
• src/physicalai/inference/model.py - validate policy_name against [a-zA-Z0-9][a-zA-Z0-9-_.]* before it is
  used to construct a filesystem path. The allowlist was derived from surveying real-world model names to ensure common cases are not broken, the positive test added.
• src/physicalai/robot/so101/calibration.py - validate SO-101 joint tick ranges: range_min < range_max ordering
  and both values within the encoder bounds.
• src/physicalai/inference/adapters/registry.py and src/physicalai/capture/transport/_publisher.py - added Semgrep/Bandit suppressions on false positives.

Why

Several inputs that cross a trust boundary (manifest JSON loaded from disk, calibration files, and user-supplied policy_name) were used in sensitive operations (filesystem path construction, recursive instantiation) without validation.

Validation

All pre-existing unit tests pass unchanged, confirming no regression in normal usage.

Added the following unit-tests (both positives and negatives):

• test_valid_policy_name_accepts_real_world_names - common real-world policy names (pi0_fast, pi0.5, smolvla, multi_task_dit, pi0-fast_v2) accepted

• test_manifest_traversal_policy_name_raises - a malformed policy name read from the manifest is rejected before any file I/O

• test_manifest_policy_name_with_slash_raises - a policy name with a path separator in the manifest is rejected

• test_explicit_traversal_policy_name_raises - a malformed policy name passed directly by the caller is rejected

• test_rejects_traversal_in_type_based_artifact - artifact path traversal is blocked for type-based component specs

• test_rejects_traversal_in_class_path_based_artifact - artifact path traversal is blocked for class-path-based component specs

• test_from_dict_rejects_inverted_range - a calibration file where min and max tick values are swapped is rejected

• test_from_dict_rejects_out_of_bounds_tick - a calibration file with a tick value outside the hardware encoder range is rejected

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds validation and safety checks to inputs crossing trust boundaries (manifest/calibration/user args) to prevent path traversal and recursion-based DoS during inference component loading and robot calibration parsing.

Changes:

• Added strict artifact path resolution within export directories and capped recursive component instantiation depth.
• Validated policy_name before using it in filesystem path construction, with unit tests for valid/invalid names.
• Validated SO-101 calibration tick ranges for ordering and encoder bounds, plus tests.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
src/physicalai/inference/component_factory.py	Adds safe artifact path resolution and recursion depth cap for component instantiation.
src/physicalai/inference/model.py	Validates policy_name (manifest + explicit arg) before constructing model paths.
src/physicalai/robot/so101/calibration.py	Validates joint tick ranges for ordering and encoder bounds.
src/physicalai/inference/adapters/registry.py	Adds Semgrep suppression comment for non-literal import false positive.
src/physicalai/capture/transport/_publisher.py	Adds Bandit suppression comments for subprocess usage.
tests/unit/inference/test_manifest.py	Adds tests for artifact traversal blocking and component depth cap.
tests/unit/inference/test_model.py	Adds tests for policy name validation (positive + negative cases).
tests/unit/robot/test_so101.py	Adds tests for SO-101 calibration tick ordering and bounds.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.