Cubase-MCP-GAJ3T.MusicGen.R.../prompts/report_8_fix.md

14 KiB

CURRENT EXECUTABLE IMPLEMENTATION — REAL NATIVE SD3 TEMPLATE SONG CODEC
Verification: 2026-10-19 local / 2026-10-20 UTC
Status: IMPLEMENTED AND VERIFIED by native saved-file structural/content
read-back. SD3 application open/play/save/reload and Cubase acceptance: UNRUN.
This supersedes the historical unavailable-layout gate, not application testing.

SCOPE

Executed the current implementation amendment atop prompts\step_8_fix.md.
Read latest plan continuation, CURRENT report_7_fix implementation and current
profiles, compiler, event/MIDI contracts, native APIs, service and tests.
Implemented formats\sdp.py, necessary strict profile/model/resource extensions,
SD3 tests/original fixtures, related resource/stdio/service regression assertions
and directly related format/architecture/profile/compatibility/README docs.

Existing Riffer implementation and working standard defaults were preserved.
No service implementation edits, native service registration, CLI template
configuration, default flips, commits, delegation or later prompt work.
The checkout already contained Step 7 and plan/prompt changes when this task
started; those pre-existing changes, including service.py, were not reverted
or misrepresented as Step 8 implementation.

READ-ONLY EVIDENCE AND FACTS

- Installed Superior Drummer 3.exe ProductVersion: 3.4.4.
- Five operator-local projects were read in place, never copied into fixtures:
  all UTF-8, PlugVersion 3.4.4, one SongTrack and a terminal NUL.
  sd3-demo.sd3p has 170 packed note triples in two clips.
  Three transcription projects each have 1,169 triples; the other project has
  no song notes. All observed packed messages have note-on status 0x90.
  All five passed actual parser + library/pad/input-channel preflight.
- Actual library location is SAMPLER/HQ/soundlib "SL-DFH".
  Native pad/alias declarations establish kickR -> 36, snareR -> 38,
  hatsTP -> 42. Only these three voices were promoted to the target subset.
  Other GM pitches were NOT promoted merely because their numbers resemble GM.
- Genuine 8332944 = 0x7f2690 independently decodes as note-on channel 0,
  pitch 38, velocity 127. Start/end triple coordinates are quarter-note beats.
- Drumtracker/Track is unrelated tracker state. Its trackstuff/tempoMap and
  all other non-song bytes are preserved, not treated as song MIDI or decoded
  into an audio-analysis feature.
- The installed Operation Manual.pdf is actually a 20,744-byte ONE-PAGE link
  sheet, not the complete SD3 manual. Local in-memory PDF inspection confirmed
  its official manual URL; that public URL still requires login. No login or
  bypass was attempted. Mapping evidence is the local setup/project pad
  declarations, NOT a fabricated manual mapping citation.
- Local provenance identifier: LOCAL-SD3-344-DFH-PADS-20261019.
  pypdf 6.21.1 was installed solely as a local evidence reader after the
  selected import failed for a missing tool. No project dependency or package
  manifest was changed for PDF reading.

No private template, factory musical sequence, manual asset or executable was
copied into source/tests, packaged as a fixture, or uploaded. Generated local
evidence retains operator kit state and stays in ignored .test-artifacts,
outside source/tests. Public fetch sent only the official manual URL.

IMPLEMENTED CODEC AND PROFILE

1. Public Python APIs:
   parse_sdp_project(raw: bytes) -> SdpProject
   inspect_sdp_template(path: Path) -> SdpTemplateInfo
   check_sdp_capability(part, profile, *, template: Path | None = None)
   require_sdp_capability(part, profile, *, template: Path | None = None)
   write_sdp(path, plan, part, profile, *, template=None) -> Path
   validate_sdp_file(path, plan, part, profile, *, template=None) -> SdpReadback
   write_sd3p / validate_sd3p_file alias the same real native implementation.
   sdp remains the request alias; actual files MUST use lowercase .sd3p.

2. Nondefault strict production profile:
   toontrack.sd3.3.4.4.core.v1
   toontrack-sd3-3.4.4-core-v1.json
   Provenance is verified_subset; version 3.4.4, library SL-DFH, core three
   pad voices only, normal/sustain only. The actual template must uniquely
   assign these pitches to matching defined pads in the evidenced xpad.
   Active sampler MIDI-learn overrides/incompatible input filtering error.
   Channel 9, linear intensity interpolation 1..127, action-free/reset-free
   scheduling and precision restrictions are explicitly generator policy,
   NOT vendor channel defaults, named sample layers or a complete SD3 map.
   Invalid promotions/version/library/voice/control/default changes fail.

3. Capability is executable, not a metadata switch:
   standard/mapping-only profiles -> sdp_profile_not_supported;
   supported target, no optional template -> sdp_template_required;
   supported target + actual valid template -> available,
     sd3_3_4_4_template_song.
   No-template checks preserve the old no-IO boundary contract.
   An explicitly supplied unreadable/missing template is OUTPUT_IO_ERROR;
   supplied corrupt/versioned/map-mismatched state is SERIALIZATION_ERROR.
   These errors are not swallowed as a MIDI fallback opportunity.
   Mutable/unchecked profile validation and wrong-part errors remain strict.

4. Actual bounded native text parser:
   Balanced blocks, quoted/escaped strings, quoted record keys, braces inside
   quoted state, Unicode, LF/CRLF, single optional terminal NUL.
   Limits: 16 MiB, depth 64, 1,000,000 tokens, 100,000 note triples,
   1..32 contiguous source tracks, numeric beat bounds <= 16,384.
   Rejects wrong/duplicate version, unbalanced/malformed state, ambiguous
   track/current-track/clip ownership, invalid tempo/meter maps, misplaced
   song music, malformed/truncated triples and invalid timing/message bytes.
   No template scripting, eval, subprocess or arbitrary executable content.

5. Actual native writer:
   Replaces ALL source SongTrack records and current-song selection with
   exactly one generated drum track/clip. Old grooves, clips, extra active
   tracks, offsets, MIDI/database references and per-song transformations
   are removed. Shared start/end, tempo/meter, velocities, drum pitches and
   channel survive as new original native events.
   Every unrelated template byte remains unchanged, including ClearTracksFirst,
   kit/sampler/mixer/editor/tracker/e-drum/metronome/host-sync state.
   The template is never written. Output uses exclusive xb on the supplied
   staged path only, flush/fsync, bounded positive size and saved validation.
   No parent/sibling/publication is created; no existing destination is
   overwritten. Template content stability is checked across the write.

6. Independent saved-content validation:
   Parses saved bytes and independently extracts status/channel, pitch and
   velocity from packed integers; it does not reserialize expected native
   text and compare the writer against itself.
   Recompiles the selected score and accounts for ALL compiled events.
   Compares exact native note multisets, common bounds, tempo/meter, neutral
   clip controls, supported structural fields and exact non-song bytes.
   Extra references/fields/tracks, mutated content or retained source-song
   semantics fail with named serialization errors.

7. Explicit supported subset boundaries:
   Only ordinary note-on triples with explicit positive-duration endpoints.
   CC/switch/status semantics, nonzero release velocity, empty music,
   score/event mismatch, wrong channels, pre-roll or tail are rejected.
   Exact finite-decimal quarter-beat coordinates, <= six fractional places;
   nonterminating 1/3 and overprecision 1/128 error rather than round.
   Different PPQs are converted exactly, not assumed to be 960.
   Absolute score start is normalized to the common score-relative zero.
   Native percussion key metadata is neutral, not a pitched-key claim.
   Failed caller-owned staging files can remain after serialization/IO failure;
   they are never a successful published artifact.

ACTUAL ORIGINAL NATIVE RESULTS — PERSISTENT AND REVALIDATED

Original score: 137.5 BPM, 7/8, one bar, 3.5 quarter beats.
Logical score interval is 8..11.5; compiled/common musical interval is 0..3.5.
Zero-based MIDI channel 9. Three original notes:

  voice        pitch velocity  start beat  end beat
  kick          36       1        0         0.125
  closed_hat    42      64        0.25      0.375
  snare         38     127        3         3.125

Actual new Events triples, independently decoded from saved native files:
  74905   0    0.125     (0x012499)
  4205209 0.25 0.375     (0x402a99)
  8332953 3    3.125     (0x7f2699)

Read-only genuine template run:
  Input: prompt-specified operator-local sd3-demo.sd3p, 75,393 bytes.
  Source SHA-256:
    06f54b02079e4b32d532ecc232b813c3fe1865799e6f5b2b21b5ddf398b0533d
  Output:
    .test-artifacts\native-sd3-evidence-v1\operator-local\original-drums.sd3p
  Actual output: 66,194 bytes; SHA-256:
    c747670d81736441928f368d8092d5363ad8e1bf3cad453af3173c1102674f88
  Source 170 events/two clips -> saved three original notes/one clip/one track.
  Saved metadata is 137.5 BPM, 7/8, start 0, end 3.5, Offset 0.
  No source song FileName/DatabasePath/KPTransform content survives inside
  the generated song. Native validation disallows those extra fields.

Exact non-song preservation:
  65,591 bytes, byte-for-byte identical before/after, SHA-256:
    10340dc5c057f2580c474ebc56277b2deae3263cee49e4674e07b8c9cc421d32
  Independently compared raw block spans:
    SAMPLER          60,637 bytes identical (includes selected kit/mixer)
    editorVars          885 bytes identical
    mixerState          146 bytes identical
    Drumtracker       1,409 bytes identical
    SongCreator          32 bytes identical
    EdrumPreset       1,024 bytes identical
    MetronomeSettings   156 bytes identical
  SAMPLER SHA-256:
    026a27006d04555942b11124b3cc8ea6803c2fca299f8efa74a0191e84b91fc5
  Drumtracker SHA-256:
    3c71b1d1079bc083a0f9249e21dec130171196dcb4e47ca48ffb0abe3ac8bd60
  Source inode/file ID 166633186213030518, size 75393,
  mtime_ns 1788614183091903900 and full content hash were unchanged.
  SyncToHost and MIDI input configuration were preserved, not silently reset.

Original minimal fixture run:
  tests\fixtures\sd3\original-template.sd3p and original-score.json contain
  newly authored original sentinel state/music, not a stripped private export.
  Output:
    .test-artifacts\native-sd3-evidence-v1\synthetic-original\original-drums.sd3p
  Actual output: 1,360 bytes; SHA-256:
    5b7e3e7e8615d58ba8b6a4e9df818caf94d0ce2eb5a4fc1dc546c6da6c3f0e6a
  Same three generated notes/tempo/meter/bounds; source fixture hash/identity
  unchanged. The minimal fixture is a codec fixture, not a redistributed kit.

All retained local files above are ignored by git. No output was promoted
into source/tests and no private input was modified.

FINAL VALIDATION — ALL PASS, NO SKIPS

Environment: existing Windows CPython 3.13.13 venv.
runTests editor integration discovered no tests for the new file; actual pytest
execution below is authoritative. Editor problems reported no errors in the
changed codec/profile/new SD3 test files.

Focused native/service/resource/package suite:
  .\.venv\Scripts\python.exe -m pytest
    tests\unit\test_sd3p.py tests\unit\test_sdp.py
    tests\unit\test_service.py tests\integration\test_sd3p_files.py
    tests\integration\test_sdp_capability.py
    tests\integration\test_articulation_midi.py
    tests\integration\test_distribution.py -q --tb=short
  RIFF_SD3_TEMPLATE was set to the prompt's genuine sd3-demo.sd3p.
  Result: 266 passed in 15.52 seconds.

Final full current regression suite:
  .\.venv\Scripts\python.exe -m pytest -q --tb=short
  Result: 1,320 passed in 41.39 seconds; 0 failed, 0 skipped.
  Prior current baseline was 1,221. Added 96 SD3 unit cases, one native
  integration case and two targeted service readiness/MIDI cases.

Final static/dependency checks:
  .\.venv\Scripts\python.exe -m ruff check src tests
    All checks passed.
  .\.venv\Scripts\python.exe -m ruff format --check src tests
    43 files already formatted.
  .\.venv\Scripts\python.exe -m mypy src tests
    Success: no issues found in 43 source files (strict project config).
  .\.venv\Scripts\python.exe -m pip check
    No broken requirements found.
  git --no-pager diff --check
    PASS.

Full/targeted package tests rebuilt real sdist/wheel offline, installed and
imported the wheel outside the checkout, verified all six packaged JSON
resources under both installed and ZIP resource loading, and ran the actual
installed stdio service. Installed-wheel native probes generated/read back
original .griff, .briff AND .sd3p files from the installed codec. The SD3 probe
uses its own original minimal template and confirms one snare note with
channel 9/pitch 38/velocity 127 and positive native size.
Owned build/install/test workspaces were cleaned; persistent evidence remains.
No dependency restore was needed for application/tests/build tools.

During development, fixture CRLF assumptions and an old exact five-resource
assertion failed; both were corrected without weakening state/resource checks.
All final suites above were rerun successfully after those corrections.

READINESS AND REMAINING LIMITS

READY: actual bounded native codec, typed optional template input, testable
read-only preflight, nondefault evidence-backed core profile, exact native
music validation, ignored original local-template evidence and package APIs.
The registry now exposes six profiles; the original three standard defaults
and operational standard-MIDI service/fallback behavior are unchanged.
Explicit MIDI with the SD3 subset also passes without template/native wiring.
Explicit SD3 native service requests without optional template fail preflight
before staging when fallback is disabled; no native service output is claimed.

NOT DONE HERE: native writer registration/service transaction integration,
optional CLI template configuration, default-profile changes, full SD3 map,
other libraries/versions/control semantics, application acceptance.
Preserved host sync/mute/routing/operator settings may affect playback or
override project tempo; structural read-back cannot certify receiver behavior.

Actual SD3 load/inspect/play/save/close/reload, acoustic pad/dynamics checks,
and Cubase routing/native/MIDI comparison remain UNRUN. No fabricated plugin
or DAW result is inferred from the real .sd3p structural/content success.