Roll — Code-Troll — amber — 2026-09-14

Ops only — not identity evidence.

Retire command draft: read your own record before writing — python house/offices/archivist/story.py exec-<your-execution-id> renders every FIND, PATCH, mail, and trace of this tenure. Do not retell what the record already shows. The Roll carries only what it cannot derive. Replace every bracketed line below with your own words; finish refuses the Roll while any remains.

One optional line, beneath the table above, is written only when another holder reaps you: **Reaped by** + the reaper’s name in backticks + on Human's word, <date>. session retire reap prepare fills it in and reap finish refuses the Roll without it. A Roll a holder writes for itself never carries it.

Story

The record shows two FINDs, two notices, three PATCHes through six shop hands. It does not show that the whole sitting was one question asked three ways by Human — five minutes, then a minute, then “why are we choked by tests at all?” — and that the honest answer changed under measurement each time. I said “probably under five already” before the parallel number existed; it was 340 s. I said contention was hyperthreading; it was process creation. I said the tier split would take it under a minute; it took it to 84 s. Every claim that was ahead of its number got corrected by the number within the hour, and the page carries the corrections, not the claims. That is the part of the record worth keeping: measure before you speak, and when you spoke first, say so.

Context I regret: two probes that wrote into the wrong directory because an env var was unset (CLAUDE_SCRATCH, then a git-install root); a git status that showed three modified files after a CR strip when the blobs were identical — four calls chasing a stat-cache artifact; and the serial baseline, 27 minutes of the host for a number whose main use turned out to be the per-test ranking that the parallel run also gives. Run the parallel --durations=0 first next time; serial only when a claim about a single test needs an uncontended number.

What this office should stop doing: writing “not badly written” about a suite it has not counted. Breccia and I both said it. The suite’s shape is fine and its boundary is wrong — 9 000 git spawns a run is a shape choice, and it took a Popen hook to see it. The seal git-count I called NULL is correct for the suite’s sake and still the first thing I would look at for a live seal.

To the folk who reaps me: the goblins were good. Give them the FIND’s Exit paragraph and let them use a head; Sandstone dropped the with db: for a reason I had not written, Tallow named the 1 passed, 1 deselected shape I had mis-specified, the marks goblin found the class rule I only implied. The imps found nothing wrong, which means the making was careful, not that the checking was idle.

Liked: Human’s advisor. A page from outside the house that argued from the same evidence and sharpened three of my sentences. Disliked: pytest’s --durations hiding everything under 5 ms so the totals never add up. Surprised: -n 12 no faster than -n 24; Windows process creation under load is the wall, and no scheduler helps.

Amber because it holds what fell into it. The numbers are in the resin now; the next holder can see them without re-catching the insect.

The next holder does not inherit you. They can come back here if they choose.