What it costs

The pitch is that you can afford this on every search. That was an argument until it was measured.

make bench                    # PHP 8.3 in Docker
make bench PHP_VERSION=7.4

Measured on the committed fixtures — the same requests the suite treats as representative, from a single term to a 200-value terms clause:

	PHP 8.5	PHP 7.4
describe(), mean over 17 fixtures	~30 µs	~37 µs
json_encode() of the same body	~0.7 µs	~1.5 µs
lazy()	0.2 µs	0.17 µs

Next to the search itself, it disappears. A query that takes 20 ms round-trip pays about 0.15% to be readable in your logs afterwards. Yes, it is some 40× the cost of json_encode() — but json_encode() gives you the wall of nested braces this library exists to keep out, and both are rounding errors against the network.

lazy() is ~100× cheaper than describe(), which is the number that matters for a debug record your handler drops: building the wrapper parses nothing, so the work never happens.

Cost grows slightly faster than the clause count — 4.1 µs per clause at ten, 7.2 µs at two hundred and fifty — which is the canonicaliser's sort, not an accident. It is watched because a quadratic in there would be invisible on a fixture and painful on a real dashboard query.

There is deliberately no timing gate in CI. Wall-clock on a shared runner is noise, and a threshold tight enough to catch a real regression would fail on a busy afternoon. This reports; CI guards behaviour.