Skip parameters the loss cannot reach

[jessegrabowski]

A parameter can be an ancestor of the loss as written and gone once the loss differentiates — an additive output bias under a second derivative. compile_train raised on it, so the only way through was to enumerate every other parameter by hand. The collected set now leaves it out; naming it in parameters= still raises, since that asserts it should train.

The stop-gradient marker walk has to stay alongside the connection-pattern check. zero_grad yields a gradient that's connected and zero, so the pattern reports it as present — swapping one for the other puts frozen parameters back in the optimizer's set, where weight decay moves them.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 90.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 96.69%. Comparing base (e9809ff) to head (da0a497).

Files with missing lines	Patch %	Lines
pytensor_ml/pytensorf/collect.py	85.71%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #88      +/-   ##
==========================================
- Coverage   96.73%   96.69%   -0.05%     
==========================================
  Files          45       45              
  Lines        1807     1814       +7     
==========================================
+ Hits         1748     1754       +6     
- Misses         59       60       +1     

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.