[danigb]

CI runs were taking about 7 minutes, and Run tests was 354 s of that 415 s. Everything else — install, formatting, build, both generated-code checks — came to 61 s together.

Measured on this PR's own CI run: 415 s → 170 s.

Where the time went

Not where it looked. Measured by re-running the suite three ways:

Variant	CPU time
jest --coverage (what CI ran)	377 s
no coverage	342 s → coverage was 8%
no type-checking	334 s → ts-jest was 2%

The remaining ~334 s was the test bodies, and most of it was not DSP. The suite made 3.5M expect() calls, 3.2M of them in the innermost loop of an exhaustive sweep. A matcher call costs ~19 µs whether it passes or fails, so the arp sweep alone spent 17 s inside jest proving the same invariant 900,000 times.

Benchmarked directly: 900k expect() in a loop is 16,760 ms; the same checks as && with one assertion after the loop is 26 ms.

What changed

Sweeps report once. Same corners, same order, same seeds, same conditions — only the reporting moved out of the hot loop. Failure messages get better, since naming the offending corner (mode, set size, octave, step) beats an expect(index).toBeGreaterThanOrEqual(0) that has lost its loop variables by the time it prints.

Negated forms are written !(a >= b) rather than a < b so that a NaN still fails, matching the matcher it replaced.

Verified by mutation: deleting the note fold in arp's dsp.ts still fails "keeps the pitch class when it folds", exactly as before. A mutation that this style misses was checked against the original tests too — they miss it as well, so nothing was narrowed.

Suite	Before	After
arp/dsp	50.7 s	3.6 s
clock/dsp	33.0 s	4.5 s
wavetable-oscillator/dsp	27.0 s	7.5 s
virtual-analog-filter/filters	46.8 s	18.6 s

Build split from test, tests sharded four ways. The tests import package sources (jest maps @synthlet/* at packages/*/src), so they never read dist/ and had no reason to wait for a build. Formatting, the build and its two checks now run beside the shards. Six shards were tried and the maximum did not move, so four it is.

--coverage dropped from CI. No coverageThreshold, no upload, no artifact: the report was written to a gitignored directory and discarded. npm run test:coverage still produces it on demand, and npm test is now the fast path locally too.

Result

This PR's CI run, all 2823 tests passing:

Job	Time
build	66 s
shard 1	67 s
shard 2	73 s
shard 3	160 s
shard 4	105 s
test (gate)	3 s
wall clock	170 s (was 415 s)

No shipping code changed — only test files, the workflow, and package.json.

What still sets the floor

packages/karplus-strong/src/dsp.test.ts is 147 s of shard 3's 160 s on its own. Jest shards by file, so that one file is now the critical path and more shards cannot help. Its cost is genuine DSP rather than matcher overhead — it renders 30 s and 60 s of audio per test to measure decay, which needs the horizon.

Getting below ~170 s means splitting that file or shortening its renders. Left alone here: it is a change to what the tests measure, not to how they report, and it deserves its own PR.

Note on branch protection

test is the required status check on main, so a matrix job of that name would have produced test (1)…test (4) and blocked every PR. test is now an aggregate job that passes only when the build and all four shards did; the name survives and protection needs no re-pointing. Confirmed green on this run.