feat(owner_eval): standalone owner eval pipeline (GSM8K/MMLU/PPL -> P…

[present42]

…rometheus)

feat(owner_eval): standalone owner eval pipeline (GSM8K / MMLU / PPL → Prometheus)

Summary

Adds connito.owner_eval, a daemon the subnet owner runs independently of any
miner or validator. Every N cycles it loads the latest merged full model and runs
a pluggable benchmark suite — GSM8K perplexity, GSM8K task accuracy, MMLU
accuracy — publishing the results as Prometheus metrics for the leaderboard
backend to scrape.

Today the only signal is validator scoring's relative (baseline_loss − val_loss)
deltas per miner; there's no independent, absolute view of how good the
collectively-trained model actually is on real benchmarks. This fills that gap.

What's added

• Evaluator registry (registry.py) — add a metric by dropping a file in
  metrics/, decorating with @register, and listing its name in config. Each
  evaluator returns dict[str, float], so one can emit several scalars.
• Three metrics (metrics/): gsm8k_ppl (reuses
  shared/evaluate.evaluate_model), gsm8k_task (greedy generation + numeric
  extraction), mmlu (per-choice log-likelihood argmax, seeded representative
  sampling across all 57 subjects).
• Cycle-gated daemon (run.py) — polls the cycle API (get_phase_from_api),
  runs when cycle_index % N == 0. --once / OWNER_EVAL_FORCE_RUN for
  one-shot; deferred heavy imports keep the gate unit-testable.
• Model source (bootstrap.py) — chain (read-only fetch of the latest
  validator HF checkpoint via load_model, no wallet keys required) or
  base (pretrained model, wallet-free canary).
• Config — EvalPipelineCfg + OwnerEvalConfig in shared/config.py
  (model_source, n_samples_per_metric, eval_interval_cycles,
  telemetry_port, sample_seed).
• Telemetry — owner_eval_metric{metric}, owner_eval_status{metric},
  owner_eval_run_info{model_revision,cycle_index},
  owner_eval_last_run_timestamp, owner_eval_loop_heartbeat_total +
  best-effort helpers in shared/telemetry.py; mycelia_owner_eval scrape job
  in observability/prometheus.yml.
• Docs — connito/owner_eval/OBSERVABILITY.md: how the backend API consumes
  the metrics (PromQL recipes, response shapes, alerts).
• Console script weightnet-owner-eval; example owner_eval.canary.yaml.