fix(gp-import): correct bass string count, lead/rhythm roles, preview note count (#601)

Four tester-reported GP-import issues, all in the converter/parse layer:

* String count (bugs 2 & 4): <tuning> is padded to 6 slots, so a 4-string
  bass, 5-string bass and 6-string guitar were byte-identical and the real
  count was lost — a 5-string bass played on 4 strings and a 4-string bass
  showed a phantom B in the editor. Record the authoritative count in a new
  <tuning stringCount=N> attribute (gp2rs._build_xml) and trim the padded
  tail back to it on read (song.parse_arrangement). All consumers already
  trust a non-6 tuning length (arrangement_string_count, the editor's
  _stringCountFor and build-time _normalize_tuning_to_count), so this fixes
  the create-mode preview AND the built sloppak with no consumer changes.

* Lead/Rhythm reversed (bug 3): guitar arrangements were named by appearance
  order (first guitar -> Lead), swapping roles for files that list Rhythm
  before Lead. Honor 'lead'/'rhythm' in the GP track name; unhinted tracks
  keep positional fallback. Applied to both convert_file's fallback (the
  editor's track_indices-without-names path) and _auto_select_gpx, with
  cross-role dedup so name-based and positional labels can't collide.

* Preview note count (bug 1): the importer's per-track count included
  tie-continuation notes, which are folded into the previous note's sustain
  and never become separate RS notes (260 shown vs 241 imported). Exclude
  tie destinations so the preview matches the imported result.

Adds regression tests for all three. Bug 5 (no stems from synced audio) is
environment-dependent (best-effort demucs backend) and not addressed here.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

lib/song.py

@@ -1090,6 +1090,20 @@ def parse_arrangement(xml_path: str) -> Arrangement:
         while el.get(f"string{i}") is not None:
             tuning.append(_int(el, f"string{i}"))
             i += 1
+        # Authoritative string count, written by the GP/RS serializer
+        # (gp2rs._build_xml). The schema pads `<tuning>` to 6 slots, which
+        # erases the 4-vs-5-vs-6-string distinction for standard tunings;
+        # when the real count was recorded, trim the padded tail so
+        # arrangement_string_count / the editor see 4 or 5 instead of 6.
+        # Absent (archive / legacy sources) → leave the 6-slot tuning as-is.
+        sc = el.get("stringCount")
+        if sc is not None:
+            try:
+                n = int(sc)
+            except (TypeError, ValueError):
+                n = 0
+            if 1 <= n <= len(tuning):
+                tuning = tuning[:n]
 
     # Capo
     capo = 0