``` Step 8 report — honest SDP capability and require boundary Date: 2026-10-19 SOFTWARE CAPABILITY BOUNDARY: PASS — implemented, persistent and verified. REAL SDP SERIALIZATION / NATIVE ACCEPTANCE: BLOCKED — no established layout/codec. REAL NATIVE ROUND TRIPS / SD3 AND CUBASE APPLICATION ACCEPTANCE: UNRUN. STEP 9 INDEPENDENT STANDARD-MIDI SERVICE/FALLBACK READINESS: YES. SCOPE AND FILES Executed only prompts\step_8.md under the amended plan's functional-MIDI continuation decision. Read the amended plan, reports 1–7 including the standard-fallback correction, current shared NativeCapability/AGP boundary, strict profile/summary/error APIs and related architecture/compatibility records. Passing AGP software does not close its separately blocked real native gate; no Step 7 fix or later prompt was executed. Created: - src\riff_groove_builder\formats\sdp.py - tests\unit\test_sdp.py - tests\integration\test_sdp_capability.py - prompts\report_8.md Updated: - docs\formats\sdp.md - docs\architecture.md - docs\compatibility.md Reused the shared NativeCapability and checked_profile without changing them. Existing AGP, profile/resource, compiler, MIDI, schema/error, package and server APIs were preserved. No manifest changes, dependency restoration/installation, commits, delegation, service wiring, later prompts or background processes. Existing unrelated worktree content and prior reports/plan were preserved. IMPLEMENTED CONTRACT formats.sdp exposes: - check_sdp_capability(part: Part, profile: InstrumentProfile) -> NativeCapability - require_sdp_capability(part: Part, profile: InstrumentProfile) -> NativeCapability Both reject non-drums requests first, defensively revalidate the entire mutable profile using checked_profile, preserve the existing SD3/drums/sdp association, and reject a valid profile belonging to another part. For every valid matching drums profile, check returns the existing immutable, frozen/slotted NativeCapability with exactly: part = "drums" format = "sdp" available = false reason = "sdp_layout_not_established" This includes production standard.gm-drums.v1, synthetic instrument mappings, and clearly test-harness-only complete mapping attestations. Mapping provenance, product/version strings and metadata cannot enable a nonexistent native codec. The current profile model rejects native_available=true; unchecked mutations are revalidated rather than treated as authorization. Require freshly checks and raises NativeFormatUnavailableError: code = NATIVE_FORMAT_UNAVAILABLE details = { "format": "sdp", "part": "drums", "reason": "sdp_layout_not_established" } Other failures remain distinct and are not fallback: - Lead, bass or invalid request part: InvalidRequestError / INVALID_REQUEST; details format=sdp, reason=unsupported_sdp_part. - Valid wrong-part profile: ProfilePartMismatchError / PROFILE_REQUIRED; details format=sdp, part=drums, reason=wrong_part. - Wrong target/format, corrupt/deeply mutated profile, invalid unchecked model or wrong model type: InvalidProfileError / SERIALIZATION_ERROR; details reason=invalid_profile_model. - Unexpected validation/programming errors propagate. Neither API accepts paths, performs filesystem IO, returns an artifact path, serializes native bytes or automatically converts to MIDI. There is no fake successful write API, parser, new registry/framework or invented native fixture. Future service code must run preflight before preparing/touching output paths. Only the named unavailable condition may inform allowed, explicit MIDI fallback; bad requests/profiles and actual serialization/IO failures must stay visible. TEST COVERAGE AND ACTUAL OBSERVATIONS 79 new unit cases: - Typed immutable/slotted unavailable result, stable fields/repeated checks, no mutation of the source profile and explicit standard-only GM identity. - Synthetic and test-harness mapping metadata cannot establish a codec. - Native metadata cannot override the model's gate. - Stable named require error with complete safe format/part/reason. - Lead/bass/invalid parts and valid wrong-part profiles have distinct failures. - Revalidation rejects wrong instrument/part/native format/flag/channel, invalid/duplicate drum notes, ambiguous voice aliases, absent voices, pitched-range misuse, articulation aliases, invalid velocity layers, deep provenance mutations, inconsistent mapping kind/version/schema, unchecked copy/construct and wrong model types. - Check-before-mutation cannot authorize require; unexpected validation errors propagate; bad-request precedence is explicit. 10 new integration cases: - Real production registry selection, not an invented mapping API. - Guarded pathlib open/mkdir/touch/read/write/stat/rename/replace/unlink, builtins.open and os.open observe zero calls during valid check/require. - Empty output remains empty; a nonexistent child directory remains absent. - Existing SDP-named operator-data sentinel retains bytes, identity, size and mtime. This sentinel is NOT a native fixture or generated SDP. - Invalid part, wrong-part profile, wrong target, deep mutation, forged native flag and unchecked model fail before guarded IO and preserve existing data. - Separate explicit drums-only and three-part descriptions use the actual resolver, registry, composer, compiler and MIDI writer/read-back APIs. Saved files begin MThd, are type 1 at 960 PPQ, have exact track names, all track endings at 3840 ticks, eight drum hits using GM notes 36/38/42 on zero-based channel 9 with positive legal velocities. Independent Mido parsing and production expected-content validation both pass. Only the actual .mid file exists; no .sdp/.agp artifact or fake path exists. These tests prove software preflight and independent standard MIDI, NOT a native SDP writer, real SD3 mapping or the service's future fallback/publication policy. EXACT VERIFICATION AND RESULTS Environment: repository .venv, Windows, CPython 3.13.13. Observed installed versions: Pydantic 2.14.1, Mido 1.3.3, pytest 9.1.1, Ruff 0.17.1, mypy 2.4.0. No other runtime/OS/application was tested. Commands ran from the repository root. Test/build subprocess scratch setup: $scratch = New-Item -ItemType Directory -Force '.test-artifacts\scratch' $env:TEMP = $scratch.FullName $env:TMP = $scratch.FullName $env:TMPDIR = $scratch.FullName $env:PIP_NO_INDEX = '1' Focused new tests: & '.\.venv\Scripts\python.exe' -m pytest tests\unit\test_sdp.py tests\integration\test_sdp_capability.py -q --tb=short PASS: 89 passed in 0.34s; 0 failed, 0 skipped. Relevant accepted regressions: & '.\.venv\Scripts\python.exe' -m pytest tests\unit\test_agp.py tests\integration\test_agp_capability.py tests\unit\test_profiles.py tests\unit\test_standard_profiles.py tests\unit\test_articulations.py tests\unit\test_midi.py tests\unit\test_package.py tests\unit\test_errors.py tests\integration\test_articulation_midi.py tests\integration\test_midi_files.py tests\integration\test_text_to_midi.py tests\integration\test_distribution.py -q --tb=short PASS: 428 passed in 18.37s; 0 failed, 0 skipped. Includes all 74 AGP capability cases, profile/standard-map/compiler mechanics, actual MIDI and independent MIDIUtil checks, text-to-MIDI, package/error checks, offline sdist/wheel build/install and installed/ZIP-wheel profile resources. Those package installations are isolated existing test procedures, not project dependency changes/restoration. No full-suite execution is claimed. Initial owned-file checks: & '.\.venv\Scripts\python.exe' -m ruff check src\riff_groove_builder\formats\sdp.py tests\unit\test_sdp.py tests\integration\test_sdp_capability.py PASS: All checks passed! & '.\.venv\Scripts\python.exe' -m ruff format --check src\riff_groove_builder\formats\sdp.py tests\unit\test_sdp.py tests\integration\test_sdp_capability.py PASS: 3 files already formatted; no reformatting needed. Final whole-project checks: & '.\.venv\Scripts\python.exe' -m ruff check . PASS: All checks passed! & '.\.venv\Scripts\python.exe' -m ruff format --check . PASS: 60 files already formatted. & '.\.venv\Scripts\python.exe' -m mypy --strict src tests PASS: Success: no issues found in 35 source files. git --no-pager diff --check PASS: no tracked-diff whitespace errors. Existing implementation is untracked; Ruff separately checks the actual Python files, not only tracked diffs. Editor problems reported no errors in the three new Python files. Editor runTests returned "No tests found in the files"; this discovery limitation is not counted as a pass or skip. The repository-venv pytest executions above provide the actual execution evidence. No pytest/lint/format/type failure occurred; no ignores or configuration relaxations were added except the existing narrow convention for the untyped Mido import in the new test. EVIDENCE LIMITATIONS AND REAL NATIVE BLOCKERS No new downloads, lawful native specifications or controlled SDP fixtures became available. The prior official SD3 manual login requirement remains an evidence limitation, not an invitation to bypass authentication or infer a layout. No credentials, vendor login, licensed-content copying or private upload occurred. docs\formats\sdp.md records the intended future _drums.sdp role and the exact evidence still needed: - Installed SD3 product/build/version, OS/plugin mode, sound library, preset, selected map/version and exact native save/export/load/import operation. - Proof the file category stores generated drum-song/groove notes, not only preset settings or unresolved external references. - Actual container/encoding/version and musical representation; compression, embedded/referenced data, library/map dependencies, structural invariants, lengths/checksums, field ownership and unknown-field behavior. - Lawful empty/one-hit/multi-hit exports and controlled tempo/meter, onset/duration, velocity and supported articulation/control changes. - Rights for redistributable fixtures or a documented operator-local procedure. - Evidence-based serializer validation, genuine fixture/version/corruption checks, followed by real generated-file import, inspect, save, close/reload in the exact SD3 setup and separately recorded Cubase/routing acceptance. Public GM facts do not establish the SD3 map or SDP layout. A preset-only container is insufficient for drum-song generation; a scope contradiction would require a decision, not a different guessed proprietary extension. Renamed MIDI and self-consistent invented JSON/XML/native encodings are not native support. No such bytes or claimed native files were produced. Real SDP serialization, native fixture/parser round trips, corrupt-native parsing, SD3 application acceptance and Cubase acceptance did not run: UNRUN, not pytest skips or passing tests. Full native acceptance stays BLOCKED. The separate AGP native/application blockers likewise remain unresolved. CLEANUP AND CONTINUATION Final relevant UUID test/build workspace count: 0. Root build directory: absent. Existing caches, venv, dist products, scratch and unrelated worktree content were preserved. No background service started. Only owned implementation/tests/docs and this single-fenced report persist. Ready for Step 9's independent standard-MIDI service/fallback implementation: YES under the amended plan, reusing both honest unavailable boundaries, explicit standard-only identity/warnings and the verified MIDI pipeline. Step 9 still owns fallback permission, error translation, staged publication, filesystem confinement and complete MCP wiring; none was implemented here. Passing capability software is NOT a finished native writer or full release. Stopped after Step 8 verification/reporting; no later work started. ```