feat(core): teaching marks fg/ch/sd — wire + GP import + sd derivation (§6.2.2)

[byrongamatos]

Part of #334

PR 1 of 3 for the teaching marks epic (feedpak 1.5.0 §6.2.2). Mirrors the bend-shape precedent (#531/#532). The three marks are display/teaching only — nothing in the scoring / NoteVerifier path reads fg/ch/sd.

What this adds

Three OPTIONAL per-note fields, all default-omitted on the wire:

• fg fret-hand finger (-1 unset, 0 thumb … 4 pinky)
• ch strum-group key (>= -1; notes sharing a value >= 0 are one gesture, pkd gives direction)
• sd scale degree (0..11 chromatic offset above the active key's tonic; author value wins, otherwise derived)

Core wire (lib/song.py)

• Note.fret_finger / strum_group / scale_degree; emit (fg/ch/sd) and decode via the existing _wire_int_optional helper. Lossless round-trip.
• _parse_note reads the GP-written fretFinger chart-XML attribute.
• Pure helpers: key_to_tonic_pc (parse a keys.json key name like "Em" → tonic pitch class), scale_degree_for_pitch, and base_open_string_midis / pitch_from_base / note_pitch_midi (tuning offsets + capo + fret → MIDI, mirroring app.js _TUNING_BASE_MIDI).

GP import → fg

• GP5 (lib/gp2rs.py): note.effect.leftHandFinger (Fingering enum) → fg, reusing the chord-path value convention; carried as a fretFinger XML attr.
• GP8/GPIF (lib/gp2rs_gpx.py): per-note <LeftFingering> (classical p-i-m-a-c letter codes, verified against real GP8 exports) → fg.

sd derivation (server.py highway_ws)

Where notes, tuning and the keys.json timeline converge: for each note/chord-note, if the author didn't set sd and a key is active at the note's time, derive it from the sounding pitch. Base hoisted out of the per-note loop.

Tested locally

• pytest tests/test_song.py tests/test_gp2rs.py tests/test_gp2rs_gpx.py → 292 passed. (Pre-existing unrelated test_diagnostics_redact.py failure on clean main is untouched / not in this run.)
• GP8 end-to-end: converting a real GP8_Test.gp emits fretFinger="2"/"1" for the M/I notes and omits the Open note.
• Independent Codex review (read-only): caught a real bug in v1 — arr.tuning is offsets from standard, not absolute MIDI — now fixed via base_open_string_midis/pitch_from_base; re-review clean.

Notes / follow-ups

• sd derivation only fires when a keys.json is present and the key name parses; missing key / unparseable name / no tuning entry → sd left unset (never fabricated).
• Renders in PR 2; authored in the editor in PR 3.

🤖 Generated with Claude Code