fix(leaderboard): record strategy hyperparameters in the run fingerprint

[ajbarea]

The bug (latent leaderboard-correctness gap)

Both DB producers built the run config with config["strategy"] = strategy_name(strat) — only the strategy name. So a strategy's hyperparameters (Krum f, FedProx mu, TrimmedMean k, MultiKrum/Bulyan f/m, …) were never recorded. config_fingerprint's docstring claims "strategy … their params … participates in the hash", but they couldn't — they weren't stored.

Consequence: Krum f=2 and f=3 share a fingerprint, so every per-fingerprint board (accuracy, wall-clock, comm-cost, pareto, robustness) silently averages different experiments into one row. It also left hyperparameter_sage (roadmapped) with no hyperparameter data to mine.

Found while picking up hyperparameter_sage — verifying the data foundation before building on it.

The fix

• strategy.strategy_params(strat) -> dict (mirrors strategy_name; dataclass fields → dict; {} for paramless FedAvg/FedMedian/ArKrum, so their fingerprints are unchanged — no churn).
• Extract mcp_app._real_run_config(strategy, **fields) from the heavy _execute_real_training, recording both the name (for the runs.strategy column + display) and strategy_params. The existing canonical-JSON config_fingerprint then resolves params with no fingerprint-logic change. Extracting it also makes the fix unit-testable without running real training.
• Only run_real_training is touched — run_demo takes a strategy string and parse_strategy can't parameterise a bare name (Krum needs f → falls back to FedAvg), so it's paramless-only: no bug, no change (YAGNI).
• Corrected the config_fingerprint docstring to state the mechanism.

No MCP tool-signature or CLI change → no surface-hash re-pin, no docs catalog churn.

Testing

• TDD: strategy_params over every strategy shape; _real_run_config records params and Krum f=2 ≠ f=3 fingerprints (the regression guard for exactly this bug).
• Whole-repo make lint (ruff + ty) green; full suite 381 passed, 9 skipped.

Research

The Sage guard-rails this unblocks (quote sample size + variance, hard-fail on too-few seeds) are statistically grounded: hyperparameter-choice variance markedly affects ML benchmark results (arXiv:2103.03098). hyperparameter_sage itself stays on the ROADMAP — its "matched runs" semantics + thresholds are a design pass worth a human eye.

[codspeed-hq]

Merging this PR will not alter performance

✅ 41 untouched benchmarks
⏩ 10 skipped benchmarks¹

---

_{Comparing feat/strategy-params-fingerprint (ac0c279) with main (eec467b)}

Footnotes

1. 10 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩