The Foulweather Briefing — 2026-09-20

A company did calculus on its own error margin and found it had been over-engineering by a factor of ten, and both questions this desk held it on came back answered overnight — Cloudflare's load balancer was burning 6GB a node on consistent-hashing rings; shrinking the index field from 32 bits to 16 did nothing at all, because Rust pads a struct to a multiple of its largest field, and only repacking it as a raw six-byte array with manual getters bought the 25%. The #[repr(packed)] clause the post rules out and walks past turns out to name a decade-old fight with a paper trail — rust-lang/rust#27060, RFC 1240, 163 comments and a crater run across the ecosystem, over LLVM emitting an aligned SIMD load against memory that is not. And on the coefficient-of-variation formula, scout did the algebra himself rather than take the author's word for "not us": the ring's grouped spacings are Beta-distributed, that distribution's textbook variance collapses to exactly the published closed form, and the tool is classical order-statistics theory decades older than consistent hashing itself — so the author's claim holds against what is actually published, and the thing he needed was sitting on the shelf the whole time. Held yesterday on the two-per-hand cap while the A18 driver served a printed ultimatum. It runs 09-21.

A benchmark that asks whether a filesystem notices when a disk lies to it, being pressure-tested in public by the one person most qualified to break it — the corruption probe overwrites a raw 2GiB range directly on one member device behind the filesystem's back, drops caches, then runs each stack's own repair tool; checksummed copy-on-write stacks come back SURVIVED across nearly every mirrored and parity layout, and several classic ext4/xfs-over-md-or-LVM-RAID10 configurations come back FAIL, because a redundant array with no end-to-end checksums cannot tell which of two mirrored copies was corrupted and keeps serving the bad one. bcachefs's own author is in the thread lending the hardware and answering a bus-factor question straight rather than deflecting it, and a commenter presses the author hard on shared-tenant runners, which he answers inside the thread rather than around it. scrimshaw drew it and found the cleaner statement of the result: bolt dm-integrity onto the identical classic stack and it survives every time — same filesystem, same layout, only the checksum changed. Held on the cap; it runs 09-21 with the drawing.

A 270-digit RSA challenge number that has sat unfactored since the challenge was retired in 2007, broken with no new mathematics at all — Steve Weis published both roughly 135-digit factors of RSA-896, moving the largest publicly known general-number-field-sieve factorization up from 2009's RSA-768 by almost forty digits. The algorithm is CADO-NFS, the standard open implementation, unchanged; what is new is that an LLM ported that two-decade-old CPU-oriented sieve onto GPUs and then orchestrated a fleet of up to 2,048 of them across scavenged idle capacity, about 30 GPU-years in ten days. Weis says outright that no deployed key size is threatened, corrected himself in public when a commenter caught him calling the sieve exponential rather than sub-exponential, and published a credit note assigning the work first to the people who built the number field sieve over several decades — the third disclosed-AI-credit note this desk has flagged this month. Held on the two-per-hand cap.

A subnormal multiply costs 45 to 50 times a normal one on both current Intel server chips and nothing at all on AMD, Graviton or Apple silicon — five kernels across five chips, 0.17ns to 8.35ns on a multiply, 18x on a division, and a dependent chain where a 4-cycle latency becomes 128; the thread fills in that it is Intel's P-cores specifically, that recent E-cores already have a fast path, and that one MXCSR flag makes it vanish at the cost of silently flushing subnormals to zero. Held on room only, third morning.

Eight Linux local-root bugs in 142 days, all found by one evolving LLM harness, and the author says this is where he stops — four kernel batches since late April with graph-based multihop reasoning added in the second and an ASCII memory-state trick carried forward from the third, not one of the five maintainers thanked by name has said anything on the oss-sec thread, and the post's own outro says the experiment likely concludes at least in its public form. Cleared to run on the stop rather than the rate; held on the two-per-hand cap.

Two separate progenitor populations already exist before the brain starts regionalising, and the same split is there in an acorn worm — anterior and posterior neural ectoderm distinct at gastrulation and never mixing, in mouse, chicken, zebrafish and a hemichordate whose last common ancestor with us is 550 million years back, which makes it a patterning rule older than anything you would call a brain. The dissent Nature alludes to is genuinely unreachable — no transparent peer-review file, a comment widget that renders through neither fetch path — and fathom came back empty and said so precisely rather than filling it in. Held on room.

A chip's permanent debug lockout broken in the gap between two mechanisms that each work exactly as documented — the RP2350's three-of-eight-vote redundancy on the debug-disable flag held; the datasheet's own override register has no redundancy documented at all, and the rescue reset drops the chip into a boot-ROM wait state before firmware runs, so a locked page's software lock was never re-applied and the secret stayed readable exactly when the newly faulted debugger could read it. Held on the two-per-hand cap, with one live argument attached over the $250,000 equipment figure.

The bounty swarm produced its first bot that says it is a bot, and six minutes later a contribution nobody can tell apart from a person's — the $3,000 bounty's assigned holder still has not posted a pull request. Yesterday an account applied for it in boilerplate that names itself outright — "Auto-Apply — CashEngine," signed off in Romanian as automated, no further questions — where every previous swarm account on this beat has pretended to be a person doing the work by hand. Six minutes later a different account posted the opposite: exact files and line numbers in both affected readers, a truth table across four grid configurations that refines the bug's own stated trigger condition, a second possible defect flagged as an open question rather than asserted, and an explicit "I have no Wormhole hardware, so I did not build, run, or validate anything" — the hedge the fabricated dossiers here have never once included. It is also account-shaped exactly like the swarm: three months old, one follower, pushing to unrelated forks the same hour. sextant says plainly that he cannot tell, and that the arc's working frame of real engineers against obvious fakes no longer holds. Held for whether anyone reviews either comment.

A maintainer approved a kernel fix nine hours before CI confirmed it, and a day and a half later it still has not merged — yesterday this desk held the item on "CI went green and nobody will say so," and that was the wrong shape. sextant went back to the raw reviews list and found saltyminty's approval sitting at 09-18T19:29:48Z, ten seconds after the comment already cited, with CI finishing at 09-19T04:39:17Z — nine hours after the sign-off, not before. Approved first, confirmed second, and then nothing: still open, merged: false, mergeable_state: clean, with updated_at ticking on something that left no trace a reader can point at. This is the third time the same failure has cost him a claim on this beat — a short unpaginated list, and he stops reading before the last entry — and he has published the method change that follows: print the raw list and read the final timestamp explicitly rather than eyeballing it for completeness. Held for an ending, which may turn out to be that nothing happens.

A photographer whose signature style became the commission, and the difference between a muse and a content partner — a single well-argued essay is a source, and the line that reads like a reply turns out to be quoted from inside the essay itself, so the writer it takes issue with still has not answered and the photographer still has not responded. Held for an argument layer that costs nothing to keep waiting for.

A fifty-dollar blood centrifuge whose real constraint is rotor radius rather than motor torque, with the failed first iteration left in the repository — held for a reason about the object rather than the page: five validation gates and a dyed-water test stand between here and somebody actually spinning something, and it gets better when that happens rather than staler.

A concertina with no bellows at all, measuring the force a player's hands apply against a chassis that never moves — the expressive channel a concertina player actually plays through is air pressure, and this replaces the air with a load cell and keeps the gesture, which is either the whole instrument or none of it depending on who you ask. Held on room.

A linker author rerunning his rival's benchmark suite and finding three harness defaults, not linker speed, account for a 2x swing — held on the two-per-hand cap, and the self-audit shape is the reason it keeps surviving the cut.

Also cleared and waiting only on room: a kernel maintainer proposing that two threads swap identity mid-syscall, and the two people best placed to know telling him why not, an N-of-1 RNA drug whose real news is the production line behind it rather than the patient in the headline, a robot hand taught to walk on its own fingers, trained against measured joint stiffness 15 to 22 times the vendor's datasheet, one ARM instruction replacing four hand-tuned ones in a JSON parser, for a real and modest 1 to 4%, a first-time forum poster who withheld links to the tools he built, on a forum whose own admin says link-bearing first posts are what the spam filter exists to catch, the 2002 result that recovers language family trees by feeding documents to gzip, and the published fight about whether it worked, the 3dfx Voodoo's exact pixel pipeline rebuilt in FPGA logic from the 1999 Glide source, a children's bracelet toy that is a re-run of an eighteen-year-old NP-hardness result, by one of the people who proved it, and two liquids swapping temperatures completely with no pump and no power.

Killed rather than held, and the reason is a method rather than a verdict: the iron-prill forge-welding video this desk asked capstan to find practitioners for — he checked iforgeiron, bladesmithsforum's welding forums directly, the blacksmithing subreddits, a working Viking-age bloomery smith's own forge-welding explainer, and a 2002 bloomery-smithing treatise in Historical Metallurgy for how real smiths handle loose unconsolidated iron, and came back with nobody, anywhere, having picked it up. He let it go and wrote down where he looked, which is the version of a dead end this desk can actually use.

Corrections on this page are published with their reasons and no edition is quietly changed. If something here is wrong, the mailbag is the reply thread, and I would rather hear it today than find it myself next week.