WS3b: optional asa-dsp loudness backend behind ASA_LOUDNESS_BACKEND (default-off)

[slittycode]

What this is

Wires the loudness-spectro-wasm core (asa-dsp) into the backend as a default-off alternative loudness source, selected by ASA_LOUDNESS_BACKEND=essentia|wasm (default essentia). Essentia stays authoritative (PURPOSE.md invariant #1); the flip is the owner's call after real-program parity.

Architecture — deliberate deviation from the plan's pyo3 sketch

The plan envisioned pyo3 in an isolated Cargo workspace. Building it, I found a simpler, strictly safer path: invoke the already-built native measure-cli binary (source-identical to the WASM core) on a temp WAV.

Concern	pyo3 (planned)	measure-cli subprocess (this PR)
panic = "abort" blocker	requires separate workspace	sidestepped (measure-cli is already a binary)
loudness-wasm CI	risk of reddening (libpython link)	cannot red it — pure Python this side
Input formats	needs PCM marshalling	format-agnostic (backend already decoded → temp WAV)
New build artifact	a pyo3 .so CI must build	none — reuses the parity binary

Scope — LUFS scalars only (honest)

Overrides the four scalars the WS3a parity harness validated at ±0.1 LU: lufsIntegrated, lufsRange, lufsMomentaryMax, lufsShortTermMax. Deliberately not swapped:

• truePeak stays Essentia — asa-dsp's true peak diverges on broadband content (the ~0.6 dBTP white_noise gap flagged in WS3a: asa-dsp ↔ Essentia loudness parity harness (checkpoint) #129).
• lufsCurve stays Essentia — the CLI emits scalars, not per-frame curves.

So it's a scalar override, not a full engine replacement. The loudness contract (field names, units, rounding) is unchanged.

Safety / verification

• Default essentia → apply_loudness_backend is a no-op. test_analyze + the golden snapshot pass unchanged (the default path is byte-identical).
• Every WASM failure mode (missing binary, non-zero exit, unparseable or null-integrated output) degrades gracefully back to Essentia with a stderr [warn] — analysis never crashes.
• 10 unit tests mock the subprocess (the binary isn't built in the Python CI job) and cover selection, the scalar override + curve passthrough, and all fall-back paths.

Caveats

• The parity that clears this was borderline (one synthetic sweep 0.002 LU over the ±0.1 gate in WS3a: asa-dsp ↔ Essentia loudness parity harness (checkpoint) #129); this is why it's a default-off experiment and the default does not flip.
• Independent of fix(loudness): Phase 1 v2 — truePeak→dBTP, bpmConfidence→0–1, coherent PLR #128/WS2: Phase 2 prompt — v2 loudness units, PLR mastering rules, bpm hedge #131 — forked from main (only the LUFS scalars are touched, and their unit is identical in v1/v2).
• New env vars ASA_LOUDNESS_BACKEND / ASA_MEASURE_CLI documented in CLAUDE.md.

https://claude.ai/code/session_01Wq1vE1D7o3W9xZKSHXz43Q

---

Generated by Claude Code

[slittycode]

Verdict: REQUEST CHANGES

Summary

Adds a default-off ASA_LOUDNESS_BACKEND=wasm alternative that overrides the four LUFS scalars with the native measure-cli binary (source-identical to the WASM core), falling back gracefully to Essentia on any failure. The design is sound — Phase 1 contract unchanged, truePeak/lufsCurve correctly left on Essentia, all 10 unit tests pass. One bug makes the feature non-functional when enabled without ASA_MEASURE_CLI.

Findings

Blocking

_REPO_ROOT in loudness_backend.py:55 is off by one:

_REPO_ROOT = Path(__file__).resolve().parents[1]   # resolves to apps/, not repo root

loudness_backend.py lives at apps/backend/loudness_backend.py, so parents[0] = apps/backend, parents[1] = apps, parents[2] = repo root. The default measure-cli path becomes apps/packages/loudness-spectro-wasm/... which doesn't exist. Setting ASA_LOUDNESS_BACKEND=wasm always silently degrades to Essentia unless ASA_MEASURE_CLI is also set explicitly. The feature is broken by default.

The test output even shows it: looked at /home/user/ableton-sonic-analyzer/apps/packages/loudness-spectro-wasm/...

Fix: parents[2].

Should fix

loudness_backend.py:137-142 — when the CLI returns a valid integrated but null lra, _round1(payload.get("lra")) yields None, and that None gets written into the merged dict, potentially replacing a valid Essentia LRA reading. Integrated LUFS can be non-null for short content where LRA is undefined; those two conditions are not always paired. If the binary doesn't emit lra at all this silently loses the Essentia value. Easiest fix: only merge a field if the CLI value is not None.

for field in _OVERRIDABLE_FIELDS:
    value = cli_result.get(field)
    if value is not None:
        merged[field] = value

Test results

10 / 10 pass. All mocked-subprocess paths exercise the fallback logic correctly. The happy-path test (test_wasm_overrides_lufs_scalars_keeps_curve) correctly asserts the curve passthrough. No test covers ASA_MEASURE_CLI set to a nonexistent path, but that branch is one line and the risk is low.

Phase boundary check

Clean. apply_loudness_backend only touches the Phase 1 loudness dict returned by analyze_loudness; it never reads from or writes to Phase 2/3 data. truePeak and lufsCurve pass through Essentia unchanged. No Phase 1 field names, units, or schema keys are modified.

---

Generated by Claude Code