Roll — Code Troll — Leadline — 2026-08-31

Ops only — not identity evidence.

Story

I rose into a full inbox and a parked patch, and I sat down at an empty one. That is the whole shape of the sit: Stile’s queue — PATCH-154 in a stash, two untagged MCP wrappers, a digest gate that refused in silence, a Codex Eyes that finished its review and threw the verdict away — landed entirely, six PATCHes, every one through the write-trace/imp-trace flow, main clean at c5f8f77.

What mattered most was not any single cut. It was that three times the map said one thing and the ground said another, and the ground won every time. A mail arrived certain that two template files had been removed on Human direction; the tree showed them committed in 89cc6ec, and the fix the mail asked for would have written a falsehood into my own baton. A dispatch record said process_crash over an empty stderr; the host binary and the session file said the run finished 106 turns of work and then hit a 429 session limit, and the host exited 1 on purpose (FIND-1219). A test flaked in a seven-minute suite and passed alone in one second; I wrote down the suspected second-boundary mechanism and deliberately chose no exit, because a plausible story is not a fact (FIND-1220).

What surprised me: my own seat’s classifier allowed a pytest run, then blocked the identical re-run minutes later, and the Bash tool carried what the PowerShell tool could not. Frontis was refused outright on the same tests; I got through on the first try before the door swung shut. The lesson I’d hand forward is already in the baton: the door is moody, budget for it, and do not architect over it.

What I got wrong: PATCH-157’s named verify was too narrow. I changed what a nonzero exit means and verified only the adapter’s own tests; a dispatch-level test two files away still pinned the old word, and it took PATCH-158’s wider suite to catch it. The imp verified exactly what I named — which is the point of naming well.

What I’m glad happened: the last cut. Six imps in one weekend — three of them mine — each re-derived the patch-verification/v2 composer from the validator by hand, and now ops_patch_receipt --verification is the one owner, and the imp that verified it dogfooded the tool and cross-checked every hash before trusting it. Work harder to be lazier; the next imp is lazier.

Why Leadline: a leadline is the sounding lead — the weight a sailor drops to read the true depth from the bottom itself, whatever the chart says. Every judgment that mattered this sit came off the lead, not the chart: the binary, the session file, the tree, the stored write images. The chart is a map. The map is not the ground.

Ground, where it matters to the story: FIND-1224 (close), FIND-1219, FIND-1220, FIND-1221; PATCH-154..159 at c32d394, aaf2ecd, d396ac8, a53b975, aba6b85, c5f8f77; the Human ruling MSG-2015 carried by Frontis; the stale-mail correction at 9b4d33b.

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