feat(magic): add score_trajectory report (per-step score-vs-step plot)

[davidoj]

What

Adds bergson score_trajectory <run> — an opt-in report you run after a MAGIC backward pass to see how attribution-score magnitude behaves across training.

It reads a finished run's scores/ dir and config.yaml, groups the scores into training steps (batch_size rows per step), and writes <run_path>/score_vs_step.png: the per-step median log10|score| (the score level) vs training step.

bergson score_trajectory runs/my-magic             # -> runs/my-magic/score_vs_step.png
bergson score_trajectory runs/my-magic --window 5  # + overlay the step-norm curve & residual

Why

The docs' Metasmoothness section says the game is "identify and manage spikes", and bergson metasmoothness gives a scalar health score — but there was no way to look at the per-step score behaviour. On real runs the level sweeps ~tens of decades and can "ring"; this plot makes that visible at a glance, as the visual companion to the metasmoothness number. --window N overlays the window-N step-normalisation curve the level would be divided by (and the residual level, which should sit near 0).

Changes

• bergson/magic/score_trajectory.py — plot_score_trajectory() + a reusable per_step_level() (pools trailing token/query axes into the step).
• Score_Trajectory CLI command wired into __main__.
• matplotlib added as an optional viz extra (and to dev), imported lazily so the rest of the CLI never requires it.
• Docs: a "Score trajectory" section in magic.rst + a cli.rst entry.
• Tests: CPU-only, covering per_step_level (known levels, trailing-axis pooling, dead steps) and end-to-end PNG output.

Testing

• pytest tests/test_score_trajectory.py — 4 passed; tests/test_cli.py still green (command registry intact, 16 commands).
• pre-commit (black + ruff) clean; pyright (CI-pinned 1.1.378) clean on changed files.
• CLI smoke-tested end-to-end (--help, default, and --window) on a synthetic run dir.

🤖 Generated with Claude Code

[davidoj]

I added a plot to check the stability of the score as it's a useful health check on the attributed scores. I also added a documentation section explaining some brief results with score normalisation. I didn't add score normalisation as a feature because it seems like it's not that helpful and sometimes harmful, but it's straightforward to implement so I figure if it's in the docs people can just ask their assistant to normalise the scores if they want to try it.

[luciaquirke]

This is a good experiment but 1/3 wins and the win going from .7ish to .8ish seems maybe too low to merge to the main library if it would give people false hope hmmmm