project
Linux syscall tracer used in the post to compare per-call cost of openat and newfstatat between Linux-side and Windows-side runs.
SQLite 3.53's ALTER COLUMN SET NOT NULL made six write calls in a developer's test, where the copy-and-rename rebuild made 69,843. The null check stays cheap only if the column is indexed, and Ubuntu's packaged 3.45.1 cannot parse the statement.
Reality
Before judging agent-written code, a dev.to author seeded a small C backup client with eight known behaviors and observed it with system-call traces, a loopback capture and a digest check to see which observer caught what.
Reality
Ed25519 tells you which key signed which bytes. A slice log from the Ranex project shows how a policy path read from the working tree let the gated side pick the rules, and what pinning the trust root to the evaluated commit cost.
Reality
A reproducible strace benchmark on a twenty-line validation task logged 752 attempted /proc/*/environ opens from Claude Code and none from Codex, which makes where an agent looks at startup a procurement question.
Reality
A dev.to write-up argues the WSL2 slowness everyone complains about is a per-crossing cost, not a tuning problem, and hands you an A/B benchmark to prove it locally.
Reality