Study 10: test frozen-random GRU reservoir transfer

## Context

Study 01 shows that a trained H=128 multisubject GRU transfers from source-training AIND mice to held-out AIND mice. It remains unclear whether that result requires learned recurrent dynamics or can be reproduced by the reservoir-computing capacity of an otherwise untrained GRU core.

## Design

- Add one opt-in wrapper flag, `model.training.freeze_gru_core`, defaulting to `false`.
- Freeze all GRU input, recurrent, and bias parameters; train only source subject embeddings and the choice readout.
- Run Study 10 at H=128, D=614, E=4 for the three Study 01 seeds using the native initialization and unchanged training/checkpoint-selection budget.
- On each held-out AIND mouse, adapt only a fresh E=4 embedding with the existing 500-step, lr=0.001 procedure.
- Compare paired per-mouse normalized likelihood against immutable Study 01 trained-GRU references. Non-inferiority requires the mouse-bootstrap lower 95% CI of reservoir minus trained GRU to exceed -0.002.

## Done when

- [ ] Wrapper freeze flag is documented, defaults off, and has tests proving the frozen GRU core is bitwise unchanged while embeddings/readout update.
- [ ] Study 10 variant, decision record, glossary, immutable Study 01 reference manifest, and launch provenance are committed.
- [ ] One GPU smoke and the three-seed H128/D614/E4 reservoir matrix complete on Beaker hub infrastructure.
- [ ] Frozen seed- and mouse-level results pass provenance/parity validation.
- [ ] A reproducible Study 10 report overlays the unconnected reservoir point on the historical Study 01 scaling curve and reports the paired bootstrap non-inferiority result.

## Notes

First pass only: no E=8, no reservoir D curve, no recurrence-disabled control, and no reservoir-specific initialization tuning. CPU-only work must not run on Beaker. No merge without explicit confirmation.