Shards leaked memory on a 2-core container: RSS climbed steadily while the JS heap stayed flat. It wasn't JavaScript.

glibc's mmap threshold is dynamic — it ratchets upward every time a large buffer is freed. So after the first few canvas/sharp image buffers came and went, subsequent large allocations started coming from the arena heap (sbrk) instead of mmap, and arena memory is never returned to the OS.

Pinning MALLOC_MMAP_THRESHOLD_ to a fixed 128KB disables the ratchet — every large allocation uses mmap, which is returned on free. MALLOC_ARENA_MAX=2 stops the container (which reports 16 cores but is limited to 2) from opening ~128 malloc arenas that fragment and never hand memory back.

Both have to be set in code before manager.spawn(), not in .env — glibc and libuv read them at process start, before dotenv ever runs.

Uploading a profile picture signed you out, and everything afterwards crawled. The two symptoms looked unrelated and had one cause.

An uploaded avatar was stored as a data URL — up to two megabytes — and handed to updateSession, which put it on the session token. That token is serialised into a cookie, and a cookie holds about four kilobytes. Auth.js does not fail on that; it splits the token across numbered chunks — authjs.session-token.0, .1, .2 — so every subsequent request carried megabytes of cookie. That was the slowness.

The sign-out was the proxy: it looked for the session cookie by exact name, so the moment the token was chunked there was no cookie by that name, and a signed-in user looked signed out.

Matching the chunk prefix fixes the symptom. The actual fix was keeping the bytes out of the token at all — the jwt callback reads the avatar from the database and puts a URL on the token, and the image is served from a route. No future caller can reintroduce it by passing an image either.

The transcripts looked fine. But on one clip the loudest part of the waveform had no words against it.

Whisper was silently dropping whichever speaker talked first. On real team radio the driver asks and the engineer answers — so we were keeping the engineer's calm reply and throwing away the driver's question. On a project that exists to read the driver's voice, every score was measuring the wrong person, and nothing errored.

vad_filter=True fixed it: voice-activity detection segments the audio before transcription instead of letting the decoder decide what counts as speech. The trade-off is that VAD trims some word starts, so a few transcripts came back slightly worse — a garbled phrase is a much smaller failure than a missing speaker, so we kept it and filtered the garbled ones out of the demo.

The join has a similar story. Lap boundaries were first derived by assuming the race started at 17:10:00 UTC. Lights-out was 17:13:00, so all 34 radio calls sat two to three laps early — and the chart still looked entirely plausible. Only loading the session telemetry and using the real LapStartDate fixed it.