Plot per-step score magnitude at the end of a MAGIC run

[luciaquirke]

Every MAGIC run that saves scores now writes <run_path>/score_vs_step.png beside them: the per-step median log10|score| against training step. A MAGIC run takes the shuffled training documents batch_size at a time, so batch s is optimizer step s and the scores group back into per-step buckets. Steps a MAGIC backward left with no score are marked with a red band rather than plotted as absence.

matplotlib is optional, as the plot extra. A plotting failure warns and lets the run finish, since the scores are already saved. A run without matplotlib prints "Hint: run pip install matplotlib for an optional trajectory plot of scores" at startup and skips the plot.

The plot is emitted from save_magic_scores, where both scoring paths converge under global_rank == 0 and the scores are still in memory. batch_size comes from the run config, so nothing has to recover it from a finished run's config.yaml. There is no CLI command and no registry entry — bergson/__main__.py, bergson/cli/commands.py, and docs/cli.rst are untouched.

Extracted from #421, which bundled this plot with a --window step-normalisation feature. That feature is left out: its own LDS experiment found it is not a robust win (one run improved, one unchanged, one clearly hurt).

Testing

pytest tests/test_score_trajectory.py — 2 tests cover the feature: per_step_level's arithmetic (known levels over trailing axes, steps with no score), and save_magic_scores writing the PNG beside the scores. pytest tests/test_magic.py -k save_magic_scores — 2 passed, for the changed signature. The missing-matplotlib and nothing-to-plot paths were checked directly. pre-commit and pyright clean; Sphinx builds at 37 warnings, matching origin/main.

🤖 Generated with Claude Code

https://claude.ai/code/session_011TfX3ZiFEKHT2W252jPJwC