Add latent novelty query strategy using OODGuard

[ktangsali]

PhysicsNeMo Pull Request

Description

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• The CHANGELOG.md is up to date with these changes.
• An issue is linked to this pull request.
• If I am implementing a new model or modifying any existing model, I have followed the Models Implementation Coding Standards.

Dependencies

Review Process

All PRs are reviewed by the PhysicsNeMo team before merging.

Depending on which files are changed, GitHub may automatically assign a maintainer for review.

We are also testing AI-based code review tools (e.g., Greptile), which may add automated comments with a confidence score.
This score reflects the AI’s assessment of merge readiness and is not a qualitative judgment of your work, nor is
it an indication that the PR will be accepted / rejected.

AI-generated feedback should be reviewed critically for usefulness.
You are not required to respond to every AI comment, but they are intended to help both authors and reviewers.
Please react to Greptile comments with 👍 or 👎 to provide feedback on their accuracy.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[greptile-apps]

Greptile Summary

This PR introduces LatentNoveltyQueryStrategy, a third acquisition strategy for the active-learning aero example that ranks unlabeled samples by their average kNN cosine distance in the encoder's geometry latent space, reusing the existing OODGuard infrastructure. It also adds the public OODGuard.score_geometry() method that exposes the raw continuous kNN distance signal needed by the strategy.

• OODGuard.score_geometry() (ood_guard.py): clean addition with proper input validation, AMP upcasting, k-clamping, and 7 new tests covering all documented error paths and edge cases.
• LatentNoveltyQueryStrategy (strategies.py): calibrates a fresh OODGuard on the labeled pool each round, scores the unlabeled pool shard-parallel across DDP ranks, and falls back to class-balanced random on cold start. The unlabeled scoring phase has a shape bug when the unlabeled pool is smaller than world-size.
• Documentation, CHANGELOG, and config are updated to reflect the new strategy and benchmark results.

Important Files Changed

Filename	Overview
physicsnemo/experimental/guardrails/embedded/ood_guard.py	Adds public score_geometry() method that returns raw per-sample kNN cosine distances without thresholding; well-structured with proper input validation, AMP upcasting, and k-clamping consistent with _check_geometry.
examples/cfd/external_aerodynamics/active_learning_aero/src/strategies.py	Adds LatentNoveltyQueryStrategy with calibrate-then-score pipeline; the unlabeled scoring phase has a tensor-shape bug when len(unlabeled) < world_size that produces (1, 0) instead of (0, 2), breaking padded_all_gather.
test/experimental/guardrails/embedded/test_ood_guard.py	Adds 7 tests for score_geometry() covering shape contract, OOD monotonicity, error cases, k-clamping, AMP inputs, and silent logging.
examples/cfd/external_aerodynamics/active_learning_aero/src/run_al.py	Wires LatentNoveltyQueryStrategy into the acquisition strategy selector; reads latent_novelty_knn_k from config with a safe default of 10.
examples/cfd/external_aerodynamics/active_learning_aero/src/conf/al_config.yaml	Adds latent_novelty_knn_k: 10 config key and updates the acquisition comment to include latent_novelty.

_{Reviews (1): Last reviewed commit: "Merge branch 'main' into ln-al" | Re-trigger Greptile}

[mnabian]

Changes to the ODD guard look good to me.

[laserkelvin]

Minor comments

[laserkelvin]

LGTM

[ktangsali]

/ok to test 62f7fed

[ktangsali]

/blossom-ci