[razedproduction]

What

Stage 8d of the porting plan: the C API's validation layer and device functions, and an unchanged upstream program built against the port's phonon library.

• Validation layer (crates/phonon-capi/src/validation.rs, wired into every module): with IPL_CONTEXTFLAGS_VALIDATION (callers of API 4.5+), each ipl* function checks its arguments and results as api_validation_layer.cpp does and logs the same warnings to the context's log callback: method name (__func__), type, the expression as the macro stringizes it, and the value as MSVC's std::to_string / %p print it. Expression strings are built with concat!, so validation allocates nothing unless it logs.
• Devices (devices.rs): Embree, OpenCL, Radeon Rays and TrueAudio Next functions fail, as in upstream built without those libraries; retain gives null, a device list has no devices.
• sample_binaural: the plan named upstream's src/itest, but it includes internal C++ headers, links the core object library and needs GLFW/ImGui/PortAudio, so it cannot build against phonon.h. Upstream's core/src/samples/sample_binaural.cpp (only phonon.h) is built unchanged against both upstream (sample_binaural_upstream) and the port's phonon.dll (sample_binaural_port, with GOLDEN_PORT_PHONON_DIR); tools/golden/run_sample.py runs either.

Verification

• capi_validation goldens: 8 sections of log messages from upstream with hand-picked invalid values (NaN kinds, infinities, %f edge cases, bad enums/flags/bools, null pointers, mismatched shapes, bad probe indices, post-call checks). Match byte for byte; the same calls without validation log nothing. Device statuses EXACT.
• capi_sample: the port-built sample's output and the same calls through the crate (on every CI platform) are within 2.9e-7 × peak of upstream's (FFT, R1). The port-built check is #[ignore] (needs the reference build; steps in tools/golden/README.md).
• fmt, clippy (both), cargo test --all pass locally on Windows.

For reviewers

• Deviations (DEVIATIONS.md): per-context log callback instead of upstream's global log; validated device factories fail instead of returning empty devices; checks that would crash upstream are skipped; context-settings warnings go to the new context's callback; only validation warnings are logged.
• Upstream quirks kept are documented as R34 (e.g. the deviation model's callback is checked against IPL_AIRABSORPTIONTYPE_CALLBACK).
• The plan's 8d row is reworded for the itest → sample_binaural substitution.

🤖 Generated with Claude Code