# Step 8: SD3 native capability assessment and safe fallback boundary ## Objective and prerequisites Read the amended plan and reports through `report_7.md`. AGP capability software was reviewed and independently passed 74 tests; real AGP native acceptance remains separately blocked with `step_7_fix.md`. Implement the corresponding honest SDP capability boundary for the later service. The official SD3 manual page required login; no layout or controlled SDP fixture has been established. Do not bypass authentication or claim a preset-only container contains generated drum-song data. ## Owned files - Create `src/riff_groove_builder/formats/sdp.py`. - Create `tests/unit/test_sdp.py`. - Create `tests/integration/test_sdp_capability.py`. - Update `docs/formats/sdp.md` and relevant architecture/compatibility status. - Write `prompts/report_8.md`. Reuse `formats.NativeCapability` and the existing AGP boundary pattern rather than introducing another registry/framework. No commits, delegation, later prompts, service wiring or fake native writer. ## Implementation requirements 1. Provide typed check/require APIs for drums/SDP, defensively revalidating profiles and rejecting bad target/part/mutated model input explicitly. 2. Valid matching profiles currently produce available=false and a stable safe layout-not-established reason. Mapping provenance/metadata cannot enable a nonexistent codec. 3. Require/check raises NativeFormatUnavailableError with part/format/reason before any path is touched. Bad requests/profiles are not native fallback. 4. No fake successful write API, renamed MIDI, invented XML/JSON, or fixture that merely agrees with an invented parser. 5. Document expected future drums artifact role, exact content/layout/version evidence needed, selected library/map concerns and actual SD3 import/save acceptance. Public MIDI facts do not establish SDP. 6. If genuine lawful specifications/fixtures become available, implement the real serializer only with evidence-based validation; otherwise report the real native gate BLOCKED, software boundary independently testable. ## Required unit and integration tests - Typed immutable unavailable drums capability and stable named require error. - Lead/bass/invalid parts, wrong target, deep mutation and unchecked-model cases fail correctly rather than being hidden as unavailable. - Standard GM fallback remains explicitly NOT an SD3-native capability. - No filesystem side effects, fake paths or overwritten sentinels. - Actual separately generated standard MIDI remains valid and contains no SDP artifact. Test against real registry/compiler/writer APIs. - Relevant AGP/profile/MIDI/package regressions remain passing. - Run focused new tests, Ruff lint/format and strict mypy. - Real native round trips/application acceptance are reported UNRUN unless actually executed against genuine data and the real installed application. ## Acceptance and report Software capability boundary can pass while real SDP serialization remains BLOCKED. Report both, preserving the independent fallback continuation decision. Do not call an unavailable capability a finished native writer. When complete, write your end-of-step report to ./prompts/report_8.md including: what was implemented, test results (pass/fail with details), any issues encountered, and whether you're ready for the next step Use exactly one plain triple-backtick fence for the entire report. Include files/APIs, exact checks/results, evidence limitations, native/app blockers and readiness for Step 9's independent standard-MIDI service/fallback path.