Add a 2D dense feature-grid consistency check (after soma seg/detection validated)

[clemsgrs]

Context

We currently guarantee pooled feature equivalence end-to-end via test_output_consistency (full tiling → PRISM encode → aggregate, compared to committed GT fixtures). We do not have an end-to-end real-weight consistency check for the 2D dense feature grid (encode_tiles_dense).

The dense code is pinned offline by the timm-oracle in tests/test_dense_extraction.py (pretrained=False: geometry, prefix-token stripping, dynamic-img-size noop). The previous real-weight spatial-registration test (test_dense_locality_gated) was removed as low-value/slow/flaky (PR #189) — it re-tested the same code paths with real weights but added no GT-pinned equivalence guard.

Proposal

Add a 2D feature-grid consistency regression check analogous to the PRISM pooled check: run dense extraction with real weights on a small fixture and compare the resulting (d, G, G) grid to committed ground-truth, so dense-path drift is caught against a golden reference (not just an oracle).

Prerequisite / gating

Do this once soma's segmentation and detection paths are validated healthy — those are the consumers that make a 2D dense-grid guarantee meaningful. Until then, the offline oracle + PRISM pooled check are sufficient.

Notes

• Mark it heavy and decide PR-inline vs the (currently disabled) heavy workflow at that time.
• Pick a small fixture / few encoders to keep it affordable on CPU; or run it on the heavy workflow.