fix(analysis): handle crash-first baselines and empty keep sets

[afurm]

Summary

• make the analysis notebook use the first non-crash run as the baseline consistently
• raise a clear error when results.tsv has no non-crash runs to analyze
• handle cases with no KEEP rows yet so the notebook does not fail on summary/plot generation

Why

The notebook previously used two different baseline definitions:

• plot cell: first non-crash run
• summary cell: first row in results.tsv

That means if the first logged experiment was a crash, the plot and summary could disagree about the baseline. The summary cell could also fail outright when there were no kept runs yet because it assumed idxmin() was always valid.

What changed

• plot cell now guards valid.empty before reading the baseline
• plot y-axis bounds now handle the no-kept-runs case safely
• summary cell now derives the baseline from the first non-crash run
• summary cell now prints n/a values instead of failing when there are no kept runs yet

Scope

This is analysis-only. No training code or experiment behavior changed.

Validation

• verified analysis.ipynb is still valid JSON
• did not run the notebook end-to-end in this environment