Test the last three: recordings, the library, and working a sound out

[ibrahimweng]

The end of the ten. samples.ts says who made a recording and what may be done with it, freesound.ts fetches them, rebuild.ts works out how a sound was made.

The credit path has a cost attached

A library assembled from Freesound is a mix of licences, and both ways of getting it wrong are expensive: a credit that should have been written and wasn't is an obligation quietly unmet; one written where none was owed pads a credits file until nobody reads it, and the ones that matter go down with it.

Both halves are tested against each other rather than separately — Freesound writes a licence as a deed URL coming back and as a name going out, and each was already right on its own when the fault the comments describe was live.

Three dead options removed

SearchOptions had four fields; one travelled. A length cap, a page number and a page size sat there unused — searchUrl never wrote them, the proxy never reads them, nothing ever passed one. How long a sound may be and how many come back are fixed server-side (duration:[0 TO 30], page_size: 24), so a caller asking for fifty per page got twenty-four and no way to tell.

A correction to what I claimed in #48

rebuild is the only thing here with no right answer available at runtime, so the only way to measure it is to render a voice the app owns and ask the search to find its way back. Twelve voices, either side of the transform fix:

	before	after
named first	58%	75%
in the three offered	83%	83%

When I fixed that transform I said it lost "every comparison between sounds that are not the same." That was an overstatement. A transform that scrambles every spectrum the same way still found the right voice more often than not — which is exactly why nothing noticed it for so long. The fix is real, and it is smaller than I made it sound.

The thresholds in the test are floors taken from the worse of those two numbers, so it fails if the search degrades and stays quiet if it improves.

Verified by breaking three things

fault	tests that failed
the CC0 check back to matching the start of a string	2
a deed URL stored unread as the licence	1
a query pasted into a URL rather than escaped into it	1

One cost worth naming

The unit suite is 62 seconds now rather than ten. All of that is the recovery measurement, which renders forty candidates against each of eight voices. It's the only check there is on whether the app can do the thing it exists for, so I kept it — but it's a real change to the local loop, and easy to trim if you'd rather.

359 unit tests (was 333), 88 browser tests, build clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ux1kydvUkLRoMbHp82ofDi

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
beat-studio	Ready	Preview	Aug 31, 2026 12:55pm