Checkpoint 7.3: an hour under load, no drift

The soak: 1.56 million timer/fetch iterations over one hour — pending handles identical at minute 1 and hour 1, 389,730 fetches with zero errors, and an RSS shape that says allocator, not leak.

The last check on the roadmap is the one only time can run: does the runtime accumulate anything it shouldn’t? One hour of soak.js — eight setTimeout(0) per iteration, a setInterval that cancels and reschedules itself every three ticks, and a loopback rt.fetch every fourth iteration — while perf/soak.py samples RSS once a minute.

The answer, two ways

The runtime’s own view: rt.pendingTimers() reported 9 at minute one and 9 at hour one — across 1,558,920 iterations and 1,565,093 interval fires. Whatever a cancel-path leak would look like (pending climbing forever), this isn’t it. The handles come back.

The OS’s view is the more interesting read. RSS went +52MB over the hour — the number that would normally get reported as “drift +52MB, investigate”. But the shape says something different: flat ~30MB for fifteen minutes, then a single step to ~88MB, then flat for thirty-five minutes — including one visible drop to 59MB and back, which is a garbage collection caught mid-sample — then a final step down to ~82MB for the last stretch. A leak is a slope: memory that climbs and never comes back. This is plateaus and a dip: an allocator sizing its heap to the working set, holding it, occasionally returning some. The pending count staying at 9 confirms it — retained work never grew; the heap just learned its own size.

389,730 fetches, zero errors, on the same loopback serve path the million-request run measured. The serve target’s RSS stayed ~24MB all hour — flat as a table.

What “kept” means

The results directory also carries the two 20-second smoke artifacts from before the real run: one where the workload itself was wrong (r.text() — rt.fetch returns {status, body}, not a Response) and 6,630 fetches all failed, and one verifying the fix. They’re evidence the smoke-test-first discipline caught a harness bug before it could burn an hour.

And with this, the roadmap is done — M0 through M7, every step with committed evidence: the run lifecycle, the module fence, the capabilities, the supervision boundary, the stdlib, a real app shipped and rolled back, and the load numbers that say what it costs.

Engineering lesson: drift is a shape, not a delta. “RSS went up” is the start of a question, not an answer — read the series, look for slopes vs steps, and check the runtime’s own retained-work counters before calling anything a leak. The worst memory bug reports are the ones that were just an allocator doing its job.