[JFMGC-54] Report pooled health topology via health_instances

[yakavaliou]

• Add JointFMClient.health_instances() to probe each configured endpoint independently and report availability, per-instance max_sample_count, summed parallel capacity, and a compact topology (e.g. 1x7000, 1x3000).
• Keep health() consensus metadata using the minimum reachable cap as the sample-batch size; endpoints remain unaware of each other.
• Surface topology in the CLI health command and document the sum-vs-min distinction in README / API reference.

>>> from jointfm_client.client import JointFMClient
>>> from dataclasses import asdict
>>> 
>>> client = JointFMClient.from_env()
>>> asdict(client.health_instances(refresh=True))
{'instances': ({'deployment_id': '6a7f76b43a393b6b724e0819', 'metadata': {'status': 'ok', 'schema_version': 'v1', 'image_version': '0.2.0', 'model_version': 'jointfm-inference:0.2.0+ckpt.chevron_i576_o288_f16_t8_h16l16_mam7_af_t3r2_cnn_k3l4_hpst_h16l2_studentt_m2cr5df8', 'checkpoint_version': 'chevron_i576_o288_f16_t8_h16l16_mam7_af_t3r2_cnn_k3l4_hpst_h16l2_studentt_m2cr5df8', 'checkpoint_path': '/models/jointfm.pt', 'device': 'cuda', 'head': 'studentt', 'decoding_strategy': 'parallel_dense', 'supported_query_modes': ('forecast',), 'supported_return_modes': ('log_prob', 'mean', 'quantiles', 'samples'), 'supported_time_index_modes': ('absolute_datetime', 'continuous_float', 'ordinal'), 'time_index_encoding': 'legacy_discrete_grid', 'max_sample_count': 5000, 'data_generation': {'sampler_type': 'fin', 'min_features': 0, 'max_features': 16, 'min_targets': 1, 'max_targets': 8, 't_input': 4.0, 't_output': 2.0, 'n_input': 576, 'n_output': 288}}, 'error': None}, {'deployment_id': '6a80af033a75b0a67a4df8ab', 'metadata': {'status': 'ok', 'schema_version': 'v1', 'image_version': '0.2.0', 'model_version': 'jointfm-inference:0.2.0+ckpt.chevron_i576_o288_f16_t8_h16l16_mam7_af_t3r2_cnn_k3l4_hpst_h16l2_studentt_m2cr5df8', 'checkpoint_version': 'chevron_i576_o288_f16_t8_h16l16_mam7_af_t3r2_cnn_k3l4_hpst_h16l2_studentt_m2cr5df8', 'checkpoint_path': '/models/jointfm.pt', 'device': 'cuda', 'head': 'studentt', 'decoding_strategy': 'parallel_dense', 'supported_query_modes': ('forecast',), 'supported_return_modes': ('log_prob', 'mean', 'quantiles', 'samples'), 'supported_time_index_modes': ('absolute_datetime', 'continuous_float', 'ordinal'), 'time_index_encoding': 'legacy_discrete_grid', 'max_sample_count': 5000, 'data_generation': {'sampler_type': 'fin', 'min_features': 0, 'max_features': 16, 'min_targets': 1, 'max_targets': 8, 't_input': 4.0, 't_output': 2.0, 'n_input': 576, 'n_output': 288}}, 'error': None}), 'max_sample_count': 10000, 'topology': ((2, 5000),), 'topology_label': '2x5000'}

New parameters in the response

• topology --> compact summary of reachable caps only, grouped by size — e.g. 1x7000, 1x3000 or 2x5000. Unavailable peers are omitted.
• max_sample_count (top-level) --> sum of reachable peers’ caps = overall parallel sample capacity (7000 + 3000 → 10000). Not the same as service.max_sample_count.

---

Note

Low Risk
Additive public API and CLI output; existing health() batch-cap and forecast batching behavior is preserved.

Overview
Adds JointFMClient.health_instances() and public InstanceHealth / HealthInstances types so callers can see per-deployment probe results when using JOINTFM_DEPLOYMENT_IDS (or a single endpoint).

Health probing is unified: health() still returns consensus HealthMetadata with max_sample_count = minimum reachable cap (sample-batch size for forecasts); health_instances() lists every configured peer (including failures), sums reachable caps as parallel capacity, and exposes topology / topology_label (e.g. 2x5000). The pool’s PoolHealthGate and probe_health() now attach per-peer outcomes; caching stores both consensus and instance rows.

The jointfm-client health CLI prints instances, topology, and top-level max_sample_count (sum, distinct from service.max_sample_count). README and API reference document the min-vs-sum distinction.

^{Reviewed by Cursor Bugbot for commit 085951d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[shackmann]

Very nice. Thank you!