fix(architectures): initialize Trompt's init_rec prompt parameter

[ChrisW09]

Part of #423 (item 6). This one is currently making CI intermittently red on main.

Problem

Trompt.__init__ creates its initial prompt representation as

self.init_rec = nn.Parameter(torch.empty(config.P, config.d_model))

and never initializes it, so the parameter keeps whatever was in the allocation. When that memory holds garbage the model emits NaN logits.

This is not theoretical — it is the cause of the flaky CI failures observed on unrelated PRs:

• tests/test_models.py::test_experimental_lss_fit_predict_evaluate[TromptLSS] on Tests (Python 3.11, windows-latest):
  ValueError: Expected parameter loc (Tensor of shape (32,)) of distribution Normal(...) to satisfy the constraint Real(), but found invalid values: tensor([nan, nan, ...])
• tests/test_models.py::test_experimental_classifier_fit_predict_evaluate[TromptClassifier] on Tests (Python 3.12, ubuntu-latest)

Both appeared on a PR touching only deeptab/metrics/classification.py — the failures are unrelated to that change and reproduce from this uninitialized parameter.

The sibling prompt embeddings in ImportanceGetter (deeptab/core/inspection.py:32-35) use the identical torch.empty pattern followed by torch.nn.init.normal_(..., std=0.01). init_rec simply missed the call.

Fix

Apply the same torch.nn.init.normal_(self.init_rec, std=0.01). As a bonus this makes Trompt seed-reproducible, which it could not be while a parameter came from uninitialized memory.

Tests

tests/test_trompt_init.py:

• Monkeypatches torch.empty to return NaN-filled tensors — simulating a garbage allocation deterministically — and asserts init_rec is finite after construction. Verified this fails on main and passes with the fix.
• A smoke test that Trompt predictions are finite after a fit.

🤖 Generated with Claude Code