Measure the overlap, not the wall clock, in the preprocess concurrency test

[adamjohnwright]

This test failed on CI for #186, a change that touches nothing it covers:

AssertionError: took 0.63s, expected roughly 0.40s
assert 0.6318183140000002 < (0.2 * 2.8)

The two assertions immediately above it — that each step starts before the other finishes — passed. Only the wall-clock budget failed.

That budget covered two 0.2s sleeps plus asyncio.run start-up, three coroutine hand-offs and whatever else a shared runner was doing, so it measured the runner about as much as the code. It was also a weaker restatement of what the interval assertions already prove.

It now measures the overlap itself, from the same perf_counter marks:

overlap = min(safety[1], language[1]) - max(safety[0], language[0])
assert overlap > DELAY / 2

Two steps that ran back to back overlap by ~0; two started together overlap by ~DELAY.

Verified: 30 consecutive runs, 0 failures. And it still catches the regression it exists for — making preprocess await the two steps sequentially instead of asyncio.gather fails it with "language detection started after safety finished".