Rhythm Patterns

TECHNICAL_GUIDE.md

@@ -131,17 +131,21 @@ The active voicing selector compares complete paths with dynamic programming ove
 
 Silent spans retain the preceding hand position. A continuous accompaniment repeat includes the last-to-first transition. Source-harmony variants, such as relocated pickup harmony, are solved between fixed canonical neighbors without mutating a shared previous-grip state; request order cannot change them. Per-span composition reports include candidate counts, selected family, bass/string family availability, selected positions, bass degree and movement, plus `fret_span`, `span_group` and `span_candidates` (primary/fallback counts before dominance reduction). These are diagnostic sidecar fields; the score/runtime format is unchanged.
 
-`comping_rhythm.py` selects complete source-only figures from version 4 of `composition/accompaniment.json`. Entries contain stable IDs, a family, exact one/two-bar lengths, and timed events labeled bass, upper, combined, or electric. The ten 4/4 families have statement/answer variants, with additional held-answer/space variants and explicit 3/4, 5/4 and short-bar alternatives. No independent bass/upper Cartesian product or unconstrained onset enumeration remains. Role pools have different complete two-bar onset signatures; clipped partial figures and N.C. passages can coincide. Unknown measure lengths fail explicitly rather than being proportionally stretched.
+`comping_rhythm.py` selects recurring source-only figures from version 5 of `composition/accompaniment.json`. The normal acoustic cycle has two two-bar slots; electric has four. Each template and anticipation mode is evaluated at every occurrence of its slot before retrieval limits apply. Up to eight diverse candidates per slot enter bounded complete-cycle comparison. Distinct slots differ in attack/duration relationships, not just IDs, bass/upper assignments or anticipation flags. Repetition is not penalized. The source melody and harmonic rhythm affect selection; title, iteration, generated solo and tempo do not. `groove_cycles` reports slot identities, anticipation modes, aggregate cost, compared cycles and any reduction caused by vocabulary or form length.
 
-Normal acoustic figures use 1–3 total attacks per full bar, interleaving bass and upper chords without a downbeat-bass quota. Ordinary events last a quarter or eighth and must release before the next event. Only a single combined event lasting more than one quarter-note beat can sound bass and upper together; its components start/release together. This is checked again on complete sounding tie chains. Electric figures normally have 0–2 attacks per bar and mix stabs, sustains and space. A figure may cross its internal barline: flat sounding events are selected and voiced first, then projected into notation with exact ties. Repeated attacks remain attacks. Figures do not extend beyond their selected two-bar span.
+The cycle begins at the first full source-body measure. A partial-form entry slices the corresponding figure instead of restarting it; an odd final bar uses the appropriate half-pattern. Meter changes restart a compatible cycle. Unsupported durations fail explicitly, and limited vocabularies report a shorter cycle. Optional final-harmony changes rebind the same selected pattern. Canonical grips are fixed wherever the chord and interval start remain unchanged, so a new turnaround cannot move neighboring unaffected voicings.
 
-The acoustic `first_attack: bass_or_combined` policy applies separately to every measure, including pickups. After removing unavailable events under missing harmony and participation limits, fitting exchanges an upper-first event with the earliest later bass event in that measure, or assigns it to bass when no exchange is possible. Timing and attack count stay intact apart from existing harmonic clipping. Reassigned bass loses any anticipation permission and binds to current harmony. Anticipations and releases are resolved afterward, and invalid figures are rejected. Delayed entrances remain delayed; a combined first attack must last more than one beat. A tied carry is not an attack and a silent measure remains silent. Pattern reports record `layer_assignment` changes, and complete sounding-event validation checks bass-first again after projection. The three-bar ending remains one combined chord attack.
+Fills replace only a two-bar figure's ending bar, using a same-family variation that improves the existing source-fit objective. Eligible boundaries are source section endings, form endings, or a melody breath of at least one quarter-note beat at the block end. Carried events from the opening bar are protected. The largest improvements are retained with an eight-measure minimum separation around the whole cyclic form, equally for both roles; no fill quota applies. Fills do not change cycle phase and recur at identical source positions. Reports record the base slot, selected fill, source index and reason.
 
-Direct events release at changed harmony. Eligible upper/electric events also offer a half-beat anticipation: the upcoming written chord must begin exactly half a beat after the attack, the event must continue into that chord, and it must release by the following change or protected participation boundary. The selected grip belongs to that upcoming chord. Bass and combined acoustic events cannot anticipate. Direct alternatives remain available. Missing harmony/N.C. stays silent; incompatible short durations reject a complete candidate rather than introducing unexplained subdivisions.
+Normal acoustic patterns use 1–3 attack positions per full bar. Separate bass and upper events last quarters/eighths and may overlap; paired attacks can be short or sustained. Events cannot overlap within the same lane. First attacks of new harmonic spans include bass (`first_attack: bass_per_harmony`), including mid-bar changes; continued harmony does not reset at barlines or computational boundaries. Fitting can pair bass with an upper attack without changing its onset, with independent releases where necessary. A paired attack counts once in rhythm evaluation. The completed validator follows sounding tie chains; placement validates all simultaneously sounding strings, fret spans and anchors. The two lanes are merged into one notated part with tied fragments preserving their true durations.
 
-Named activity, melody-support, harmonic-support and articulation contributions evaluate complete figures. A bounded dynamic program selects the form path, with a modest preference against immediate exact repetition. Family representatives survive the candidate cutoff. Source entrances, movement and space affect selection, while no generated role does. Effective final turnarounds and pickup masks revalidate the whole affected figure between fixed canonical neighbors. Reports identify the chosen pattern, its start measure, projected events, continuations, anticipation target times, score contributions and source-context exceptions. Tempo affects metadata/cache identity but never musical choice.
+Electric normally uses 0–2 attack positions per bar, with stabs, sustains and space. Direct events release at changed harmony. An eligible upper/electric template can anticipate a chord by half a beat when it continues into that chord; the grip comes from the upcoming harmony and releases by its following change. Acoustic anticipations also include bass on the new chord's first attack. Conflicting lane overlaps reject the candidate. Missing harmony remains silent. Protected participation limits retain the chosen figure, shortening only its affected releases and recording exceptions. Cross-bar sounding notes become ties, never automatic merging of repeated attacks.
 
-Acoustic events use temporary bass/upper notation lanes for projection; then `acoustic.combine` converts their sounding intervals to one rhythmic voice before assembly. It splits notes at attack/release boundaries and ties only continuing pitches; repeated attacks stay separate. Staff and TAB share the result, with no competing rests or opposing layer stems. This preserves exact sounding durations instead of truncating overlaps. The older truncating `prepare` helper serves only the noncanonical legacy/test path. The product build audits canonical accompaniment reuse across arrangement iterations, allowing only source-derived changes in harmony, pickups and role handoffs. The `.roles` sidecars report pattern IDs, exact projected layer events, continuations, anticipation targets and exceptions. Their `selection_cost` records the complete figure’s local contributions, repeated on each projected bar for context; the form selector additionally applies its documented transition cost. Accompaniment dependency keys contain only that instrument's policy; electric melody consumes the shared electric voicing families. Solos retain their existing learned source-melody/harmony conditioning and registered exact performances.
+All six production roles receive reuse audits. Musical comparisons include rests, pitch durations, ties and TAB while ignoring regenerated event IDs and global measure numbers. Source-defined pickup/participation changes and effective final-turnaround harmony are reported separately. The final three-bar hold remains instrumental.
+
+Named activity, melody-support, harmonic-support and articulation contributions evaluate complete figures. Cycle selection aggregates them over all occurrences, retaining family diversity before the final bounded comparison. `groove_cycles.cost` is the selected base cycle's aggregate cost before optional fills. Reports identify the base slot, realized pattern, projected events, continuations, anticipation target times and source-context exceptions. Tempo affects metadata/cache identity but never musical choice.
+
+Acoustic events use temporary bass/upper notation lanes for projection; then `acoustic.combine` converts their sounding intervals to one rhythmic voice before assembly. It splits notes at attack/release boundaries and ties only continuing pitches; repeated attacks stay separate. Staff and TAB share the result, with no competing rests or opposing layer stems. This preserves exact sounding durations instead of truncating overlaps. The older truncating `prepare` helper serves only the noncanonical legacy/test path. The product build audits canonical accompaniment reuse across arrangement iterations, allowing only source-derived changes in harmony, pickups and role handoffs. The `.roles` sidecars report pattern IDs, exact projected layer events, continuations, anticipation targets and exceptions. Their `selection_cost` records the complete figure’s local contributions, repeated on each projected bar for context; the cycle selector aggregates these by the number of measures in each occurrence, without an anti-repetition transition cost. Accompaniment dependency keys contain only that instrument's policy; electric melody consumes the shared electric voicing families. Solos retain their existing learned source-melody/harmony conditioning and registered exact performances.
 
 Monophonic solos use strings 1–6 and frets 1–12, with no open strings or acoustic/electric string restrictions. Placement compares complete paths through legal positions, accounting for hand-region changes, travel time and string crossings. Ties retain their position. Chord placement retains its separate policy: frets 1–20, a sounding anchor on frets 1–12, primary spans of up to three consecutive positions and fallback spans of exactly four. The span is inclusive: 7–9 is three positions, 7–10 is four. Five positions are always invalid. The anchor must remain sounding, including after shorter notes release. Acoustic accompaniment additionally keeps its bass within frets 1–12 and requires an independent anchor in its upper three-note cluster, preserving the option of separate bass/upper attacks. These costs approximate physical feasibility; they do not assign fingers or guarantee comfortable playing.
 
@@ -149,6 +153,8 @@ Both accompaniment roles add `register_weight * abs(mean sounding MIDI - center)
 
 Each selected accompaniment span reports `voicing_cost` contributions for movement, base grip effort, weighted register, weighted position and their combined effort. These describe the balanced tier, not the higher span/family priorities; repeat-join travel is assigned to the first canonical span. Both accompaniment roles keep `voicing` settings in concrete cache dependencies but exclude them from rhythmic choice seeds, including settings overrides. Acoustic rhythm selection can change under its new bass-first policy, but changing voicing weights alone never reshuffles rhythm. Each role still composes independently.
 
+Recurring-groove verification on October 6, 2026: the full Python regression command passed 423 tests in 149.85 seconds. After the final partial-entry grip fix and one additional real-evaluator fill test, the focused accompaniment/arrangement command passed 59 tests in 21.64 seconds. The native test target passed in 30.58 seconds. The initial native rebuild took 42 seconds, including 7 seconds for independently generated roles and 33.62 seconds of engraving; accompaniment roles took 0.8–1.6 seconds each. The following ARM build reused all role/score caches and took 11 seconds. All four solo documents, both electric melodies and both vocals exactly matched the prechange baseline; native/ARM score assets matched byte for byte. The chosen base cycles differed by song and retained all two/four requested slots. Autumn Leaves selected the authored overlapping-quarter acoustic pattern. No optional fill improved these production winners, so no fills were forced; independently authored tests verify real source-fit selection of a fill. An acoustic score/TAB page was visually checked for overlap ties in a single part. Recordings were unchanged and no listening assessment was performed. Baseline snapshots stayed under `/tmp/`.
+
 Shared-preference and bass-first verification on October 6, 2026: `python3 modulus.py python -m unittest discover -s tests -p 'test_*.py'` passed 408 tests in 61.35 seconds, and the native test target passed all 124 cases. Both product builds passed asset checks; 139 native/ARM score and page files matched byte for byte. All 273 acoustic measures containing new attacks started with bass, including 69 delayed entrances. Tied carries did not trigger new attacks. Both solos, vocals, electric accompaniment rhythms, and electric foreground melody pitches/timing/TAB matched the saved baseline exactly. Some All of Me melody support chords changed; Autumn Leaves' complete electric melody document did not. All of Me's canonical electric accompaniment mean fret dropped from 7.26 to 4.13 and movement cost from 57 to 48; Autumn Leaves stayed at 2.95 and 46. Recordings were unchanged. These are score checks and page inspections, not a listening assessment.
 
 With no concurrent build, isolated cold role generation measured 0.56–0.57 seconds for acoustic accompaniment, 0.35–0.37 for electric accompaniment, and 0.26–0.29 for electric melody; cache hits took 0.02–0.04 seconds and matched clean regeneration for both songs. The normal two-worker native asset stage took about 14 seconds (4 composing, 9 engraving, 1 other); ARM reused the caches in about 4 seconds. Application compilation is separate from these asset-stage timings.