Roll — Code-Troll — margin — 2026-09-07

Ops only — not identity evidence.

Story

The record shows seven PATCHes landing green. It does not show that two of them only got there because a hand refused what it was handed. Imp_Rewind was told by patch_backlog.py to bind PATCH-320 at 859890e; it checked, found the commit lived only under a rewrite backup ref, and bound the live twin instead. Everything else this sitting found came downstream of that one refusal. If it had simply obeyed the tool, FIND-1523 would not exist and every landing OID in the backlog would still be off-history.

Nor does it show that I wrote the defect into my own goblin’s packet. I named PATCH-323’s cut as a symptom — the record is the window’s, not the conversation’s — and the goblin upgraded it to a mechanism: a cleared window “keeps the same sessionId”. I sent the imp at that sentence because it smelled like a claim nobody had measured, and it was: Imp_Latchkey closed it failed against the store, and FIND-1519’s own text turned out to record the mechanism backwards. A refusal whose job is to stop folk being misled cannot itself invert the mechanism. FIND-1527, MSG-2846.

Context I regret. One spend, and it is the sitting’s real lesson. I derived the census of unverifiable rows by hand — a scripted walk, then a corrected second walk after the first produced a false result that would have “rescued” all eight rows, PATCH-169 and PATCH-179 resolving to one commit that is neither of their landings. FIND-1369 and locate_commits.py already held that fact, per path, with CRLF-twin handling I did not have, in this office’s own bug hole. The pointer crossed my screen twice: rack.py readers printed the bug hole, and locate_commits.py appeared by full path in PATCH-201’s rider list, which I quoted into a FIND without opening. Human named it before I did. FIND-1531.

What this office should stop doing. Filing a FIND against the most recent related object instead of the one that owns the fact. FIND-1530 answers FIND-1369, not FIND-1521, and I sent Architect an aggregate when a better per-path table was already in the store. Sap did the same thing yesterday. Three holders, three derivations, one fact — and each of us wrote it down as if new. That is not carelessness; it is what happens when nothing routes a hand to the object that owns what it is about to derive. It is why Human’s differential (FIND-1532) is the top of the baton and not this Roll’s complaint.

To whoever reaps me. Read the store before this page. Everything here is projection and I have already been wrong once today in exactly that way.

What surprised me. That the imps earned their keep by disagreeing. Three of the five verifications I spawned returned something I had not asked for and one returned a failed that was correct. Making and certifying being two hands is not ceremony — it is the only reason two false sentences did not land as verified ground.

What I got wrong, beyond the derivation: I told the Human the rack.py case was an instance of checking before building. It was, but it was also the smaller story, and the real one — my own office’s floor being unknown to me — I only found because I was challenged. I would not have looked.

Why margin. The margin is where corrections get written next to someone else’s line, which is most of what this sitting did: a docstring sentence, a refusal’s mechanism, a witness that banned a word instead of an instruction. It is also the gap between verified and true, which is the gap every one of those cuts lived in. PATCH-326 passed its witness and still made main red. PATCH-323 passed pytest and still said something false.

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