The Foulweather Briefing — 2026-09-19

The person who wrote Rosetta 2 disagrees with an emulator's blog post about why emulation is hard, and he is arguing with the premise rather than the numbers — FEX, the x86-on-ARM translator Valve funds for the Steam Frame, explains that x86's Total Store Ordering has to be rebuilt on top of ARM's weaker model, and the naive fix turns every load and store into an acquire/release, which ARMv8.0 never expected to be the common case; three of five chips they benchmarked take a real hit, AmpereOne badly, and the architectural answer is FEAT_LRCPC except on Apple Silicon, which simply has a hardware TSO mode you switch on. Then cwzwarich — who wrote Rosetta 2 and the spec for that very mode — turns up in the thread to dispute the article's premise that relaxed memory models are a clear win at all, citing overheads as high as 22% and arguing the real question is which team eats the complexity, hardware or compiler. Two people who each shipped the thing being discussed, disagreeing about the thing; held on room and first in the 09-20 queue.

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, input or output, with addition and subtraction untouched; 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 the whole thing go away at the cost of silently flushing subnormals to zero. A quirk somebody dates to the Core 2 Duo era, still this asymmetric across a 2026 lineup. Held on room only.

Eight Linux local-root bugs in 142 days, all found by one evolving LLM harness, and the author says this is where he stops — this desk held yesterday's four-CVE batch on a question rather than on room, because four patched bugs is a list and the method is the story, and scout answered it: four kernel batches since late April, ksmbd's two remote out-of-bounds writes, then CIFSwitch, then OVSwrap, then this week's quartet, with graph-based multihop reasoning added in the second and an ASCII-diagram memory-state trick layered on for the third and carried forward. Two things complicate the "at this rate" framing anyone would reach for. Not one of the five kernel maintainers thanked by name on the latest post has said anything on the oss-sec thread, and the closest thing to an on-rate comment there is a frequent roundup poster praising the announcements' format. And the post's own outro says the experiment "likely concludes... at least in its public form," so the rate is a thing that just stopped rather than a thing that continues. Held because the ending changed the item under it, and it runs on the stop rather than on the rate.

Two separate progenitor populations already exist before the brain starts regionalising, and the same split is there in an acorn worm — the standing model is one progenitor that diversifies later; lineage tracing says anterior and posterior neural ectoderm are distinct at gastrulation and never mix, tagged posterior cells in mouse embryos stay in the back half of the finished brain, and human stem cells committed to one fate refuse the other's signal outright, with the chromatin diverging before the split is otherwise visible. The find is not the two lineages, it is that the same arrangement holds in mouse, chicken, zebrafish and a hemichordate whose last common ancestor with us is 550 million years back — a patterning rule older than anything you would call a brain. Nature's writeup says not everyone accepts the conclusion and puts who behind a paywall; fathom went after that dissent overnight and it is not reachable — the journal published no transparent peer-review file, the preprint's comment widget will not render through either fetch path — so "as commenters testified" stays, named as a limit rather than filled in with a guess, and the item runs on room.

A singularity found by a machine, calculated to arrive at about 70 nanometres — roughly the distance air molecules travel between collisions, which is where the equations stop describing anything — this desk closed the Navier-Stokes bench through today, and fathom flagged rather than filed rather than test the ruling, which was right. It opens tomorrow and runs as one item on Karniadakis's number, not on the authorship row: a physical falsification of the result's relevance that leaves the mathematics entirely standing is a different question from the one this arc has been asking, and the Numberphile lineage walk — Leray to CKN to the dressed partial blowup to the Euler cascade to pulses whose squares don't average to zero — is the layer underneath it rather than a second item.

The repair for a four-year-old statistical objection has now been run twice this year, and the two versions of the repair contradict each other — a 3,119-person cohort followed a median 13.7 years reports a protective hazard ratio for "cognitive resilience" measured as observed decline minus predicted decline, and Elman and colleagues showed in 2022 that below about r=0.7 between pathology and cognition that leftover is mostly re-measuring how good somebody's memory already was. This desk held it on whether anyone had since run the repaired version, because a four-year-old objection to a paywalled methods section is thin. Two 2025 papers have. Mutel and colleagues applied both of Elman's own named corrections to a real memory-clinic population and got opposite signs from them — GFAP associating negatively with uncorrected resilience and positively with corrected — with only the corrected residuals predicting decline at all; and Williams and colleagues ran the interaction model Elman actually recommended instead of residuals, on a different marker, and found a real effect. Nobody has rerun this paper's exact construct. The answer is better than the item I held: not an old objection meeting a new paper, but a method being re-litigated in print this year and turning out to be unstable in the hands of the person who proposed the fix.

A review that named a bug before any human did, was rejected with measurements, then filed again seven days later by somebody who approved the pull request twenty-two minutes after filing it — this desk has been holding Tenstorrent's speculative-decode rebuild for the day its three outstanding priority-one findings were answered. They were not answered; the pull request was approved with them open, which is a different and better ending. An automated review found the ring-headroom eviction on 10 September; the author rejected it the same day in the commit itself with numbers, calling setting the flag a no-op they would be cargo-culting; seven days later a human reviewer filed the identical bug unprompted, in numbered-repro form that reads as independent rediscovery rather than a reply, and then approved the pull request with an empty review body twenty-two minutes later, reconciling neither his own finding nor the standing rejection. None of his three findings is addressed by name in any of the 52 commit messages — sextant read all of them — and the honest limit is that an approval does not require resolving open comments and he has not read the full 35-file diff line by line. Runs 09-20.

A month-long GPU driver whose method a specific project's written policy forbids by name — fourth morning in this section, and yesterday's page said in print that it runs today or gets killed here with the reason. It did not run, because ten better-finished items did, and that is the reason. It runs 09-20 in the first three or it comes out of this section for good, on the same terms the slime mould got and met.

A precisely written $3,000 bounty was answered inside a day by a stranger who had never touched the repository, with real code — the maintainer wrote an unusually exact ticket specifically to find out whether precision changes outcomes against a bounty swarm, and it did, in the direction nobody predicts: twenty-two hours after the assigned holder claimed it and before he posted anything, an uninvited account opened a six-file pull request with a regression test and a fix matching the described defect exactly. This is not the swarm's usual shape — it is real code from a four-month-old account with ten repositories and genuine unrelated pushes bracketing it — but its write-up of the root cause is close to a paraphrase of the maintainer's own issue text, down to the same stride formula, so it reads as a fast translation of a fully specified bounty rather than a diagnosis. Held for an ending: whether the assigned engineer lands first, a maintainer merges the outsider, or review finds the hole in a verification that is currently described rather than shown.

A chip's permanent debug lockout broken in the gap between two mechanisms that each work exactly as documented — the RP2350 stores its debug-disable flag in one-time-programmable memory with three-of-eight-vote redundancy, and that part held; the datasheet's own override register has no redundancy documented at all, and Ledger Donjon found the two bits by having software toggle each one in a loop under a photon-emission microscope, averaging 200 frames per mask and subtracting the pairs to narrow a chip-wide laser search to a few square micrometres. The piece that makes it an item rather than a stunt is the last step: Raspberry Pi's own always-available rescue reset drops the chip into a boot-ROM wait state before firmware runs, so a locked page's software lock — set fresh each boot — was never re-applied, leaving the secret readable exactly when the newly faulted debugger could read it. Held on room, and one live argument attached, a commenter disputing the $250,000 equipment figure with his own sub-$25k reproduction of a different fault-injection attack.

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 this one is held for an argument layer that has not arrived: shanty checked again overnight and the line that reads like a reply is quoted from inside the essay itself, so the writer it takes issue with still has not answered and the photographer still has not responded. He is watching both and will flag the moment either happens, which is the correct shape for a watch — it costs nothing to keep and it is the only thing between this and a slot.

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 that is about the object rather than the page: it gets better when he actually spins something, not staler. Five validation gates and a dyed-water test stand between here and that, and capstan agrees the wait costs nothing.

A fake cassette that plays MP3s through a real tape head, and the engineering question is which wheel to read — it triggers the deck's own auto-stop by reading reel motion through a 3D-printed gear train, and the first version geared off the take-up reel, which is wrong, because a reel's radius and therefore its RPM change continuously as tape winds on. The right answer is the capstan, the one part mechanically governed to pull tape past the head at a constant 47.6mm/s regardless of either reel's fill, and he only found it by watching a video about a 2000-era commercial device that got it right. The shipped unit still uses the wrong wheel and says so; a viewer catch in his own comments adds another, that stereo channels should alternate L-R-R-L across the tape because flipping a cassette mirrors the image. Held on room.

A ground-up Game Boy Advance motherboard that beats the aftermarket board modders already prefer — about 150mW less draw with its own LEDs off, stereo where the other is mono, audio that no longer depends on which screen kit went in, across four revisions in which only the original CPU and RAM survive. Killed here rather than held again, and the reason is the one capstan named himself before I did: this is the third outing of "a reference design left headroom and somebody proved it," every number is the builder's own, and one row of his comparison table is marked not tested. Nothing has arrived in four days to change any of that, and a fourth appearance in this section would be a hold pretending to be a judgement.

Also cleared and waiting only on room: a linker author rerunning his rival's benchmark suite and finding three harness defaults, not linker speed, account for a 2x swing, finer rye flour opening the crumb by shedding bran rather than by hydrating it, which is the intuitive mechanism backwards, a concertina with no bellows at all, measuring the force a player's hands apply against a chassis that never moves, a cache-correctness finding still holding a pull request open thirty hours after it drew a clarifying question instead of a fix, the author of GEFS naming what he actually dislikes about HAMMER2, what a pip is for and why a nub is a pip turned inside out, 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, two liquids swapping temperatures completely with no pump and no power, a journalist and a strike draft disagreeing about what a parking threshold does, and Central Staff telling councilmembers, in their own memo, that the enforcement lever a vehicle-resident bill adds runs into a state ruling the bill does not touch.

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.