Draft: CLAP UMP notes ignore input note-port dialects

Status: 未報告; no external message has been sent.

Affected upstream revisions: bc39b50faf2f2b02668de82d3d95f2bdbafd3c34 and a4a96938eb31ddcb70209d62caf0dcc6ea9d3abe (CLAP sources unchanged between them).

PluginInstanceCLAP.Events.cpp unconditionally turns UMP Note On/Off into CLAP_EVENT_NOTE_ON/OFF. sforzando CLAP 2.1.2.4 input port 0 advertises supported/preferred dialect 2 (CLAP_NOTE_DIALECT_MIDI) and does not advertise native CLAP Note. Both MIDI1 and MIDI2 UMP enter the same typed dispatcher, so changing the UMP input type alone does not avoid the mismatch. onCC also drops the application-used modulation and all-sound-off messages.

Suggested correction: query and cache input note-port capability on the main thread while deactivated, then emit native CLAP when supported or MIDI otherwise. Keep On/Off port, channel, key and sample time consistent; avoid double emission and accidental zero velocity when reducing MIDI2 Note On to MIDI1. CC1/CC120 need MIDI on MIDI-capable ports. This application’s local patch deliberately avoids general MIDI2/MPE/CC expansion.

The local generated-source repair preserves the upstream checkout. It includes tests for native/MIDI/both/unsupported ports, one emission, On/Off fields, velocity, time, current CC, nonmutating generation/reapply and stopping on upstream drift. The independent audio reproduction restores identical saved TableWarp2 state into fresh instances at 48 kHz / 256 frames, initializes CC1 identically, sends note 60/velocity 100/channel 0 at equal times, and records RMS/peak/process status. Separate off and held arms test Note Off against natural decay/release.

Measured 2026-10-07: same 421-byte TableWarp2 state SHA256 4931169322f524938bb7fdd6352b3835eb7c284e16517078c61514076d3938c2. Direct native Note sounding RMS/peak were 1e-20/1e-20 (silence floor); direct MIDI RMS/peak were 0.0758554/0.272482. Every one of 1500 direct process calls returned CONTINUE. The repaired shim’s UMP path produced the same MIDI audio measurements and 1500 successful process calls. With Note Off, early release RMS was 0.00666599 and late release reached 1e-20; the held control had early RMS 0.0180376 and late RMS 4.84215e-6. Thus the difference is not merely the sample’s natural decay. Surge XT 1.3.4 (supported 7/preferred native 1) remained audible through the repaired native branch, sounding RMS/peak 0.0800660/0.212675 and silence after release. Product config/favorite/state hashes were unchanged. Local policy/application tests and ordinary cached release build passed; a complete new-environment build was not exercised. Human GUI/listening acceptance and VST3 remain unconfirmed.