[Feature] Bilinear probing

[Xmaster6y]

What does this PR do?

Key insights about the PR.

Linked Issues

• Closes #?
• #?

Checklist

• I have read the CONTRIBUTING guide.
• I have added tests for my changes if needed.
• I have updated the documentation if needed.

---

Summary by cubic

Adds bilinear probing to train classifiers on paired layer representations and analyze cross-layer interactions. Refactors probing into a torch-first, modular API with ProbeManager/BilinearProbeManager, updates docs/notebooks, and expands tests; removes the sklearn manager.

• New Features

  • BilinearProbeManager for layer pairs with a key_pattern regex; BilinearProbe caches activations and validates pairs via before_all/after_all.
  • Torch estimators: BilinearEstimator, LowRankBilinearEstimator, LinearEstimator, plus MeanDifferenceClassifier; optional verbose fit logs.
  • New bilinear probing notebook and docs card; GPT-2 + IMDb example.

• Refactors

  • Probing moved into tdhook.latent.probing (context, estimators, managers); exports updated; sklearn manager removed.
  • Notebooks now use ProbeManager (linear-probing and torchrl-ppo); fixed r2 callback signature in torchrl-ppo.

^{Written for commit d0477ae. Summary will update on new commits.}

[codecov]

Codecov Report

❌ Patch coverage is 97.95322% with 7 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.19%. Comparing base (b7c78fb) to head (d0477ae).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/tdhook/latent/probing/context.py	93.47%	3 Missing ⚠️
src/tdhook/latent/probing/managers.py	98.28%	3 Missing ⚠️
src/tdhook/latent/probing/estimators.py	99.13%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #41      +/-   ##
==========================================
+ Coverage   96.81%   97.19%   +0.38%     
==========================================
  Files          36       39       +3     
  Lines        2226     2426     +200     
==========================================
+ Hits         2155     2358     +203     
+ Misses         71       68       -3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[cubic-dev-ai]

4 issues found across 11 files

Confidence score: 3/5

• Potential runtime errors in src/tdhook/latent/probing/estimators.py when bias=False adds None to a tensor or when a dataset has a single class leading to divide-by-zero/NaNs.
• src/tdhook/latent/probing/context.py may raise a TypeError during hook registration if list types are passed into MultiHookManager without converting to tuples.
• Score reflects multiple medium-severity issues that could cause failures or invalid outputs at runtime.
• Pay close attention to src/tdhook/latent/probing/estimators.py and src/tdhook/latent/probing/context.py - guard against runtime errors in edge cases and hook registration.

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/tdhook/latent/probing/context.py">

<violation number="1" location="src/tdhook/latent/probing/context.py:42">
P2: Convert `classes_to_hook`/`classes_to_skip` lists to tuples before passing them to `MultiHookManager`; otherwise passing a list will raise a `TypeError` in the `isinstance` checks during hook registration.</violation>
</file>

<file name="src/tdhook/latent/probing/estimators.py">

<violation number="1" location="src/tdhook/latent/probing/estimators.py:36">
P2: Guard against datasets with only one class; otherwise mean computation divides by zero and yields NaNs.</violation>

<violation number="2" location="src/tdhook/latent/probing/estimators.py:44">
P2: Avoid dividing by a zero coefficient norm; guard or skip normalization when the norm is zero.</violation>

<violation number="3" location="src/tdhook/latent/probing/estimators.py:172">
P2: Handle bias=False explicitly; adding None to a tensor raises a runtime error.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

1 issue found across 2 files (changes from recent commits).

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="docs/source/notebooks/tutorials/torchrl-ppo.ipynb">

<violation number="1" location="docs/source/notebooks/tutorials/torchrl-ppo.ipynb:267">
P2: The R² metric lambda uses predictions as the first argument, but `ProbeManager` passes `(predictions, labels)` and `r2_score` expects `y_true` first. This swaps `y_true`/`y_pred` and produces incorrect R² scores.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

1 issue found across 1 file (changes from recent commits).

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="tests/latent/test_probing.py">

<violation number="1" location="tests/latent/test_probing.py:197">
P2: This test codifies a crash for bias=False instead of asserting correct bias-free behavior. Consider adjusting the estimator to handle bias=None (e.g., add 0) and update the test to expect a valid output shape; otherwise the test will hide a real bug.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}