namir-app: the RT harness is not wired to the real installed audio callbacks

[ErwanLegrand]

crates/namir-app/src/rt_harness.rs and src/stream.rs:560-640

audio_section has exactly one caller in the crate (audio_io/convert.rs:420). stream.rs -- which holds the callbacks actually installed into cpal -- has four tests that drive the real output callback through FakeBackend, and none of them run under the harness.

convert.rs's RT test claims "the closures the converters hold are the RT-safe shape crate::stream actually installs", but they are stand-ins; the real one allocates. Wrapping output_cb(&mut out) in audio_section(...) would catch both the callback allocation and the input-scratch realloc immediately.

Fix: run stream.rs's existing callback tests inside audio_section.

[ErwanLegrand]

Fixed. The premise was partly stale by the time #145 opened: 608fdde (M14) already wrapped stream.rs's real callbacks in rt_harness::audio_section — the_audio_callbacks_this_module_builds_allocate_nothing, now crates/namir-app/src/stream.rs:942, and it found the two allocations this report predicted (build_output's per-chunk Vec<&mut [f32]>, and build_input's mono_scratch realloc).

What was left was the integer-format path, closed by 6644aeb: the_real_stream_callbacks_run_allocation_free_inside_both_converters (crates/namir-app/src/audio_io/convert.rs:456) takes the real callbacks out of crate::stream::fake_duplex_setup, wraps them in the i32 and I24 converters a device that will not take f32 invokes, and drives both under audio_section (convert.rs:493-497). audio_section is no longer a one-caller function.

---

Generated by Claude Code