Restore granular feedback behind a show_detail parameter

[peterbjohnson]

Closes #24.

Based on 22-organise-feedback-messages (#23) so the diff shows only this change; GitHub will retarget it to main when #23 merges.

Summary vs detail, not current vs old. polished_feedback_message → summary_feedback, generate_feedback_message → detail_feedback, and the two blocks in feedback_messages.py are relabelled to match. build_feedback composes them: the summary is always shown, the detail is appended only when show_detail is on.

SHOW_DETAIL = False sits with the other module-level defaults and is teacher-configurable via params["show_detail"], like the thresholds.

The caveat. detail_caveat_message introduces the detail section. It is worded to hold for MIDI submissions too — those aren't transcribed at all, so a flat "this came from a transcription" would be inaccurate about half the time:

The per-note comments below are generated automatically and may not be accurate for every note, especially if your recording was transcribed from audio.

Duration direction bug. event_level_feedback stored duration_abs_diff = abs(...), so direction = "longer" if duration_abs_diff > 0 was always true — a note held too short was reported as too long. Rather than make the existing key signed, which would have silently inverted duration_correct (the threshold compares duration_relative_diff <= threshold, and a signed value would let badly short notes pass), the fix adds a separate duration_signed_diff and leaves the threshold logic untouched. Two tests cover both directions.

Part 1 dropped from detail_feedback. It restated the summary in the older voice. The overview_* constants went with it.

Tests. Six new, in a section 11: detail off by default, appended below the summary when on, introduced by the caveat, reachable through params, and the two duration-direction cases. 78 pass. Section titles and the caveat are asserted by constant, following 6121639.

Notebook. Phase1-1.5_summary.ipynb cells 3/40/42 are updated to the new names. Outputs are not re-run, so the printed detail output in that notebook still shows the old Part 1 overview.

🤖 Generated with Claude Code

[ada-jz7125]

the notebook may need to be rerun later, I realised that text assertions are fixed in PR27

[ada-jz7125]

this notebook may need to be rerun later to make sure the output is up-to-date