Remove XML Solo Gen

TECHNICAL_GUIDE.md

@@ -193,7 +193,7 @@ Files are copied into `data/songs` and `data/training` without overwriting diffe
 
 Production input traces distinguish authored reuse from learned prediction. Caches include source, role, relevant data and code; authored cache keys include the selected XML's content and convention. Build and package inputs include the model, manifest and authored repertoire. Changes to another solo do not change an already registered part's notes. Updating the pooled model can legitimately change predictions for any unregistered part. CMake defaults to two composition processes (`VAMP_COMPOSITION_JOBS`, 1–16) and caches each role under `build/build_common/composition_cache`. Direct generator runs use `--composition-jobs` and opt into persistent caching with `--composition-cache`; single-role builds use `--cache-dir`.
 
-The generated `solo_xml/*-solo1.xml` and `*-solo2.xml` exports use local measure numbers, preserve pickups, and store whole-arrangement positions in `manifest.solo`. Exact reuse means complete sounding pitches, onsets, releases, repeated attacks and silence agree. XML whitespace, note splitting, measure labels and exporter-specific annotation containers need not be identical. Optional `./modulus.py python -m tests.solo_compare` compares these completed exports; it never feeds a production decision. Its `--generated-dir` is the directory containing `manifest.solo`, and `--reference`/`--form-start` use the manifest's `SONG/ROLE` keys. `--require-exact` fails if any requested reference differs in complete sounding events or rests. Training and authored reuse are the explicitly permitted dependencies on the supplied music.
+Solos are published inside the arranged score JSON and engraved pages, without separate solo MusicXML exports or comparison manifests. Exact-repertoire tests compare generated sounding pitches, onsets, releases, repeated attacks, measure timing and silence directly with the authored XMLs. Tied notation fragments are merged into sounding events for this check, and generated TAB is checked against the guitar constraints. Training and authored reuse remain the explicitly permitted dependencies on the supplied music.
 
 Migration inventory:
 
@@ -215,7 +215,7 @@ Voicing-only verification on October 6, 2026, before the rhythm update: the suit
 
 Rhythm verification on October 6, 2026: `unittest discover -s tests -p 'test_*.py'` passed all 375 tests in 83.57 seconds. Coverage includes attack ranges, bass downbeats, mixed durations, harmonic-change budgets, overlapping voices, tied joins, 3/4 and 5/4, melody-conditioned rhythm costs, tempo invariance, canonical reuse, and serial/parallel role identity. A synthetic learned-model test also confirms that changing original melody rhythm or upcoming harmony changes solo duration probabilities with a fixed seed. Both songs' solo, electric melody and vocal role documents remain exactly equal to the saved pre-rhythm baseline; no model retraining was performed.
 
-The ordinary six-score rebuild with `--check --export-solo-xml --composition-jobs 2 --jobs 2` passed in 13.39 seconds: 4.44 seconds generating parts and 8.48 seconds engraving. All role composition results were newly generated; the warm engraving namespace reused two unchanged vocal scores. The subsequent `cmake --build build/build_x86_debug --target vamp_generated_assets` passed with all role/score caches reused. Actual selected acoustic upper rhythms use one or two attacks in these songs; three-attack alternatives are available and tested but are not forced. All of Me electric has three zero-attack source/harmony contexts; Autumn Leaves electric uses one or two attacks. Neither song required a density exception. Reuse audits checked 103/62 acoustic/electric repeated bars for Autumn Leaves and 102/62 for All of Me, with only source-defined pickups and final-harmony exceptions. Sample acoustic notation/TAB pages were visually inspected. Supplied audio recordings remain unchanged; no ear-based audition of the new accompaniment was performed. Disposable logs, profiles and comparisons are under `/tmp/`, outside product assets.
+The ordinary six-score rebuild with `--check --composition-jobs 2 --jobs 2` passed in 13.39 seconds: 4.44 seconds generating parts and 8.48 seconds engraving. All role composition results were newly generated; the warm engraving namespace reused two unchanged vocal scores. The subsequent `cmake --build build/build_x86_debug --target vamp_generated_assets` passed with all role/score caches reused. Actual selected acoustic upper rhythms use one or two attacks in these songs; three-attack alternatives are available and tested but are not forced. All of Me electric has three zero-attack source/harmony contexts; Autumn Leaves electric uses one or two attacks. Neither song required a density exception. Reuse audits checked 103/62 acoustic/electric repeated bars for Autumn Leaves and 102/62 for All of Me, with only source-defined pickups and final-harmony exceptions. Sample acoustic notation/TAB pages were visually inspected. Supplied audio recordings remain unchanged; no ear-based audition of the new accompaniment was performed. Disposable logs, profiles and comparisons are under `/tmp/`, outside product assets.
 
 Chord-family and unified-acoustic verification on October 6, 2026: `unittest discover -s tests -p 'test_*.py'` passed 386 tests in 81.67 seconds. New cases cover 35 chord qualities in three keys for both instruments, discarded bass variants and explicit unsupported combinations, the six suggested acoustic grips, highest feasible electric melody strings, scoped policy dependencies, and one-voice tie preservation including repeated attacks and cross-bar ties. Quarter/eighth acoustic attack durations and available consecutive-eighth patterns are tested. The acoustic sounding events in both complete arrangements contain only quarter/eighth durations, even where notation splits them into ties.
 
@@ -223,9 +223,9 @@ The full checked rebuild with two composition and two engraving workers took 13.
 
 Four-fret fallback verification on October 6, 2026: `unittest discover -s tests -p 'test_*.py'` passed 392 tests in 85.65 seconds. This includes primary-tier precedence over family, arbitrary movement savings and melody-string height; a complete held-note path that needs an earlier four-fret option; tied-fragment cost invariance; and anchor release checks. The first full run exposed one obsolete C5 rejection expectation, now replaced by a test verifying its legal four-fret voicings contain only C/G. Coverage tests check 35 qualities in 12 keys for each instrument: 418/420 acoustic contexts and 420/420 electric contexts have a legal grip. Acoustic Bbm9/Bm9 remain explicitly unsupported; this is measured coverage, not a guarantee for every written extension or slash chord.
 
-The checked six-score rebuild with `--composition-jobs 2 --jobs 2 --export-solo-xml` passed in 9.21 seconds (4.58 seconds composing all 12 roles, 3.90 seconds processing six cached engravings). All composition caches missed after the code change. Every resulting role document and accompaniment rhythm report exactly matches the saved baseline: current production chords already have primary grips, so none needed fallback substitutions. New sidecar span diagnostics identify the tiers and available candidates. The subsequent `cmake --build build/build_x86_debug --target vamp_generated_assets` passed with all role caches reused and identical per-role musical hashes. Acoustic and electric score/TAB pages were visually checked; supplied recordings are unchanged. Formatter and diff checks passed. Temporary verification files remain under `/tmp/`.
+The checked six-score rebuild with `--composition-jobs 2 --jobs 2` passed in 9.21 seconds (4.58 seconds composing all 12 roles, 3.90 seconds processing six cached engravings). All composition caches missed after the code change. Every resulting role document and accompaniment rhythm report exactly matches the saved baseline: current production chords already have primary grips, so none needed fallback substitutions. New sidecar span diagnostics identify the tiers and available candidates. The subsequent `cmake --build build/build_x86_debug --target vamp_generated_assets` passed with all role caches reused and identical per-role musical hashes. Acoustic and electric score/TAB pages were visually checked; supplied recordings are unchanged. Formatter and diff checks passed. Temporary verification files remain under `/tmp/`.
 
-The optional exact comparison confirmed all four exports: 137/141 attacks for All of Me and 121/115 for Autumn Leaves, with no pitch, onset, release or rest differences. All four now pass contextual harmony validation as well. Removing 69 explanatory annotations left the complete event streams and learned musical count tables identical to the pre-edit baseline; only model provenance hashes changed. This verifies authored reuse rather than learned reproduction. Commands used the ordinary asset generator with `--export-solo-xml --check` and `python -m tests.solo_compare` with four explicit `--reference` pairs, `--reference-pitch guitar-written` and `--require-exact`. No ear-based audition or claim of transfer quality was made.
+Earlier exact-repertoire verification confirmed all four authored solos: 137/141 attacks for All of Me and 121/115 for Autumn Leaves, with no pitch, onset, release or rest differences. All four also passed contextual harmony validation. Removing 69 explanatory annotations left the complete event streams and learned musical count tables identical to the pre-edit baseline; only model provenance hashes changed. This verifies authored reuse rather than learned reproduction. These checks now compare generated score data directly with the training inputs instead of producing standalone XML exports. No ear-based audition or claim of transfer quality was made.
 
 Forced prediction, without reading the training XMLs, completed all four source/role combinations with valid harmony. These were composition-only runs on the training-source songs, not held-out transfer tests. Their notes intentionally differ from registered playback and were not published as replacements. The expanded treatment candidates initially took 99.03–120.33 seconds per role with one worker and no concrete-result cache. Removing repeated validation of identical resolution endpoints reduced the isolated All of Me acoustic check from 104.30 to 60.13 seconds, with an identical complete output hash. Six additional contexts retained all 32,566 ordered action candidates exactly. These expanded prediction choices remain slower than the earlier 21.81–27.16-second decoder; registered playback is unaffected. Training the pooled 2,860-context model took 0.57 seconds. Prediction remains the slower path, chiefly candidate construction and event-likelihood evaluation; further profiling is needed before expanding its choices again.