Cross-species comparison / transfer

Explore cross-species comparison and transfer of the foundation model.

## Scientific question
Does the frozen GRU core learned from the AIND mouse cohort provide a useful prior for next-choice prediction in independent two-arm-bandit datasets and species?

For target subject (i), initialize a new embedding from the source-mouse mean, optionally fit only that row on an adaptation subset, and score untouched target trials. Fit subject-level Q-learning on the identical adaptation observations and score the identical test trials.

## Execution plan
1. [x] [Wrapper #91: canonical external-bandit loader and split manifests](https://github.com/AllenNeuralDynamics/aind-dynamic-foraging-bfm-wrapper/issues/91)
2. [x] [Wrapper #92: paired GRU and Q-learning target adaptation](https://github.com/AllenNeuralDynamics/aind-dynamic-foraging-bfm-wrapper/issues/92)
3. [x] [Dispatcher #127: pinned adapters for starter open datasets](https://github.com/AllenNeuralDynamics/aind-dynamic-foraging-bfm-dispatcher/issues/127)
4. [ ] [Dispatcher #126: full Study 09 benchmark and reproducible report](https://github.com/AllenNeuralDynamics/aind-dynamic-foraging-bfm-dispatcher/issues/126)
5. [x] [Dispatcher #131: Grossman author-selected meta-learning baseline](https://github.com/AllenNeuralDynamics/aind-dynamic-foraging-bfm-dispatcher/issues/131)
6. [x] [Dispatcher #132: Chen author-selected RLCK baseline](https://github.com/AllenNeuralDynamics/aind-dynamic-foraging-bfm-dispatcher/issues/132)
7. [x] [Dispatcher #133: Zid author-selected history-kernel baselines](https://github.com/AllenNeuralDynamics/aind-dynamic-foraging-bfm-dispatcher/issues/133)

## Evaluation contract
- Target test membership is frozen before fitting and identical across models and adaptation budgets.
- Grossman/Chen primary curve: K ∈ {0, 1, 2, 4, 8, half}, with a common eligible subject cohort.
- Zid within-session evaluation uses the complete manifest prefix and suffix; it has no K condition.
- GRU adaptation updates only newly appended subject embeddings; core, head, and source embeddings stay frozen.
- Q-learning parameters use only the target adaptation subset.
- Primary metric is paired per-subject test log-likelihood difference in bits/trial; geometric-mean likelihood remains for dashboard compatibility.
- Test likelihood never selects source checkpoints, fine-tuning steps, learning rates, or baseline family.
- Multi-session and single-session-prefix targets preserve true recurrent/Q-state semantics.

## Initial dataset panel
Grossman mouse, Chen mouse, and schedule-matched Zid human are the primary benchmark. Shin rat, Tang macaque, Beron mouse, baited tasks, and continuous-reward tasks are optional distribution-shift extensions after the primary result.