Stop the master clipping, and test the chain that hid it

[ibrahimweng]

The mixer and the master were the last untested part. The render tests pass chain settings through without ever asking what the chain does with them, so three bands of EQ, a tilt, a send, a room, the level, the loudness and the limiter were all exercised as far as "something came out".

Fifteen tests measured on the way out, and they found something.

The bug

The limiter was being handed a copy of the mix that had already been clamped to plus or minus one — scaled went through applyMaster, which clamps as a last resort before a file is written. So however loud the mix really was, the limiter saw a peak of one and asked for the single decibel that takes one down to the ceiling. Everything above that was then flattened by the clamp.

A mix peaking at two and a half times full scale, at three targets:

target	samples flat against the top	reported reduction
-14 LUFS	4	"1.00 dB"
-9 LUFS	628	"1.00 dB"
-6 LUFS	2160	"1.00 dB"

Exactly 1.00 every time, because that is what takes a clamped one down to 0.891. That number is what gave it away. Quiet targets were always fine, which is why nothing had noticed.

Scaling for analysis no longer clamps. The same three targets now come back at the ceiling exactly, with no flat samples and reductions of 2.46, 11.71 and 17.44 dB — and the report's loudness matches what the samples actually measure, which matters because the interface prints that number after every export.

The rest of the fifteen

• the level turns the whole thing down, and silences it at zero
• each of the three bands moves its own part of the spectrum and leaves the others where they were
• cutting a band takes it away rather than adding elsewhere
• the tone tilts the top against the bottom, which is what makes it a tilt rather than one control on one band
• the send is what puts a tail after a sound, and does not change the sound itself
• a cathedral rings for longer than a booth
• the master lands on the loudness it was asked for, reports the same number the samples say, and reports honestly that it never had to work when it did not

Spectrum is measured by running a render back through the same kind of filter the mixer is built from, since the question is only ever "more or less than the other one".

Checking

Every test checked by putting its fault back: the clamp restored fails two, ignoring the level fails two, ignoring the send fails two, dropping the tilt fails one.

251 unit tests (was 236), 77 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 30, 2026 11:35pm