The default no test was holding
Thirteen tests covered the write pool, each setting its lane count by hand. Set the default to one lane and all thirteen pass while the product goes serial.
Several screens in the back office save many rows at once: ten products given a new VAT rate, every printout type for a kitchen, a page of suggested par levels applied in one go. They do it through a small pool that keeps up to four writes in flight and starts the next as each one lands. We drove it in a browser against ten real rows on a live-shaped venue and watched the trace: four writes start together, the rest enter as lanes free, done in 135 milliseconds. The same ten rows written one after another, after reverting the single loop that pools them, took 193. Beside the database that difference is small on purpose, because almost all of a twenty millisecond write there is work rather than waiting. From the UK, where a round trip is ninety milliseconds, the serial version pays ten waits and the pooled one pays three.
Writing that paragraph is what found the gap. The pool has a default lane count of four, used by all seven screens that call it. Thirteen tests covered the pool. Every one of them passed a lane count explicitly, because each test was about some other behaviour and wanted a known number.
Change the default and watch nothing fail
So we changed the default from four to one and ran the suite. Thirteen passes. Every screen in the product had just gone serial and the traces above had become a description of code that no longer existed, and the tests had nothing to say about it. A default is a decision nobody writes down at the call site, which makes it the one value no caller-focused test will ever check.
Two tests were added. One calls the pool with no lane count at all and asserts that four writes start before any has finished. The other asserts the number itself, so a change to it is a change somebody has to make on purpose and explain. Then the proof the usual way: revert the default again, and exactly those two fail by name while the thirteen stay green. That is how we know they were the only thing standing between the product and a quiet slowdown.
The same shape elsewhere
We keep a list of the ways our own checks have failed us, and this one was already on it: a guard that stopped being true when a default changed. It shows up wherever a value is set in one place and relied on in many. A timeout nobody passes. A page size nobody passes. The lane count of a write pool. The test that protects such a value is not the one that uses it; it is the one that deliberately does not.