test(samples): end-to-end audio-content assertions for audition output

[slittycode]

What this is

Follow-up to #45. Adds end-to-end audio-content test coverage for the audition-sample feature — the layer that was missing.

Why

#45's tests verified the audition pipeline in segments:

• the plan — test_sample_theory.py asserts correct MIDI numbers for a key
• the primitives — test_sample_synthesis.py (one note → right frequency), test_sample_drums.py (synth_kick → right fundamental)
• the orchestrated output exists — test_sample_generation.py checks the WAVs are written, non-empty, and have correct metadata/citations

But nothing loaded the orchestrated WAV files back off disk and checked the audio itself reflected the Phase 1 input. test_sample_generation.py asserts the kick's label says "55 Hz" — it never asserted the kick audio was at 55 Hz. The chain was tested link by link but never end-to-end as sound.

What this adds

apps/backend/tests/test_sample_audio_content.py — 6 tests that run the real orchestrator, load the WAVs it wrote, FFT them, and assert on spectral content:

Test	Verifies
test_c_major_chord_audio_contains_the_triad	The C-major chord WAV carries real energy at C4/E4/G4, and the weakest chord tone is >20× louder than a non-chord tone (C#4) — the audio is in the key, not loud noise
test_bass_root_audio_sits_on_the_tonic	The bass WAV's dominant frequency is C2 (~65.4 Hz) within 4 Hz
test_different_keys_produce_different_audio	The load-bearing test. Generates C major and F# minor; proves C-natural is >20× louder in the C-major render than the F#-minor render, and F# is >20× louder the other way. The render genuinely tracks the measured key — it's not a canned clip
test_orchestrated_kick_lands_on_measured_fundamental	The kick WAV the endpoint serves peaks at the measured kickDetail.fundamentalHz (extends test_sample_drums, which only tested synth_kick directly)
test_kick_fundamental_tracks_the_measurement	A 45 Hz vs 95 Hz measurement yield audibly different kicks, each on its own fundamental
test_melody_follows_scale_degree_hints	Hints [1,5] in C major render C5 then G5, in that order, ascending

Scope / honest caveats

• Runs against the sine-additive fallback (prefer_fluidsynth=False) so assertions are deterministic and don't depend on a system soundfont — same reason CI can't assume FluidSynth is installed. The FluidSynth path is structurally exercised by the same orchestrator code but not asserted on.
• The bar is "in the right key / on the right fundamental" — not "sounds like the source track." That's the deliberate scope from the design doc: audition validates the measurement chain, not Ableton timbre.
• Snare/hat aren't asserted here — Phase 1 has no measured ground truth for them, so there's nothing to assert against beyond "well-formed" (already covered in test_sample_drums.py).

All 6 tests pass locally. Pure addition — one new file, no changes to existing code.

https://claude.ai/code/session_015L67xTQdgRuhCRpJurJWYd

---

Generated by Claude Code

[slittycode]

Verdict: APPROVE

Summary

Pure test addition (one new file, no production code touched) that closes a real coverage gap: the audition pipeline was tested link-by-link but never end-to-end as audio. All six tests are methodologically sound. Frontend CI passes; backend CI was still in-progress at review time — the code is clean and I'd be surprised if it fails.

Findings

Worth considering (not blocking):

• test_melody_follows_scale_degree_hints — the timing comment says "note 0 spans ~0–3.4 s" but at 120 BPM with 2 bars/note, note 0 spans 0–4 s. The sampling windows (0.5–3.0 s and 4.5–7.0 s) are correct and safely inside both notes, so the test works fine; the comment is just slightly off.

• test_orchestrated_kick_lands_on_measured_fundamental with delta=20.0 is generous (72 ± 20 Hz), but it's consistent with the delta=20.0 tolerance already established in test_sample_drums.py and justified by the pitch-envelope settling behavior called out in the comment.

Nothing else. The Hann windowing is the right call for inter-bin leakage; the 20× energy ratio is a meaningful threshold (not just > 0); prefer_fluidsynth=False correctly isolates the tests from system soundfont state; the generate_samples call signature matches the actual implementation exactly.

Test results

Frontend: pass. Backend CI in-progress at review time; no issues visible in the code.

Phase boundary check

Clean. _phase1() constructs a mock input; generate_samples consumes it. No Phase 1 data is mutated or re-derived anywhere in the new file.

---

Generated by Claude Code