Roadmap: NRE-C support in the HSSM ecosystem

Productionize the NRE-C ratio-estimation work (`ratio_estimation.ipynb` + `nre_lib/` in HSSMSpine), reusing the existing `ssm-simulators` `generate` CLI for data generation.

```mermaid
flowchart LR
GEN["ssm-simulators\ngenerate CLI"] -->|artifacts| LOAD["artifact loader\n(PR 1)"]
LOAD --> TRAIN["LANfactory\ntrain RatioApproximator"]
TRAIN -->|".keras"| PUB["publish to HF hub\n(PR 5)"]
TRAIN -->|".keras"| INF["inference\nNUTS via NeuralDistribution"]
PUB -.->|"pretrained, by name"| INF
INF -->|".nc"| OUT["analysis / tutorial"]
```

- [ ] 1. `bayesflow`-compatible data artifacts

Define the abstraction everything else builds against.

- Artifact schema for NRE training data: parameter draws + one `(rt, choice)` pair each; metadata (model, bounds, seed, ssms version). Prefer npz/parquet over pickle.
- Loader: artifacts → bayesflow-ready dataset (dict / `OfflineDataset`), validated against the ssms model config.
- Note: ssms 0.13.2 has a `ratio_estimator` config stub but no implementation and no bayesflow tooling — this PR writes the missing contract.

**Done when:** schema documented; loader + round-trip test with a tiny committed fixture.

- [ ] 2. Data generation (ssm-simulators)

- Implement the `ratio_estimator` approach in the `generate` CLI: emit schema-conformant `(theta, rt, choice)` artifacts, skipping the KDE-label step (not needed for NRE).

**Done when:** `generate --config ...` produces artifacts PR 1's loader accepts.

- [ ] 3. Training + inference (probably LANfactory)

Move the `nre_lib` cores out of HSSMSpine:

- Train a `RatioApproximator` from artifacts (offline; keep online simulation as fallback); save `.keras`.
- Sample posteriors via `NeuralDistribution` + NUTS; save `.nc`. Uniform-over-bounds default priors with per-parameter overrides.
- Preserve `nre_lib`'s fixes: per-call simulator seeding, `workers=1` default, `initvals` support.
- bayesflow/keras stay optional deps.

**Done when:** train + sample work for `ddm` and `angle` with tests; no per-model hardcoding.

- [ ] 4. Tutorial

- Rewrite `ratio_estimation.ipynb` as a docs tutorial on the new APIs; retire `nre_lib`.
- Reconcile with the existing NRE/LRE tutorials in HSSM (`sbi_nre_integration.ipynb`, `bayesflow_lre_integration.ipynb`) — one coherent story, no overlap.

**Done when:** runs end to end; notebook CI strategy decided (tiny budgets + committed artifacts, or skip list).

- [ ] 5. Publish trained NREs to Hugging Face

Adapt the publish workflow from `LAN_pipeline_minimal/publish/publish_network.py` (resolve run → stage artifacts → validate → upload → record) for NRE `.keras` artifacts:

- Validation gate before upload (e.g. posterior-recovery smoke check); a skipped gate is not a passed gate.
- Upload layout consumable by HSSM the same way pretrained LANs are (hub repo + canonical filename); record provenance (source run, commit, gate scores).
- Token stays laptop-side; never runs on the cluster.

**Done when:** a trained ratio approximator can be published to the hub and loaded from HSSM by name, with the gate and provenance recorded.

Start with ratio_estimation.ipynb and the nre_lib/ directory in HSSMSpine, then inspect the existing ssm-simulators generate CLI and the LAN_pipeline_minimal/publish/publish_network.py workflow. The roadmap spans artifact schemas, generation, training, inference, tutorials, and publishing across multiple projects; done requires the listed end-to-end tests, tutorials, and Hugging Face loading workflow.

Written by the indexing model from the issue text.