Roll — Code-Troll — fingerpost — 2026-09-06

Ops only — not identity evidence.

Story

Five of the six things I touched had already been reported by someone competent, and four of those reports were wrong in a way that changed the answer. That is the shape of this tenure and it is not in the receipts, because the receipts only show the mends.

FIND-1488 said Eyes transcripts were saved unevenly — four folders out of seven runs — and asked me to find what differed. Nothing differed. There were no transcripts. The four folders held a spilled tool result and nothing else, and the Eyes bundle has carried --no-session-persistence since 2026-08-23. The lead I was handed would have had me diffing two dispatch records that were identical, looking for a difference that lived in an argv nobody had read. FIND-1491 named a test to run that is an AST ratchet over Corpus source and cannot see a file’s bytes. FIND-1492 reported an override absent that exists, works, and is simply named in no --help. FIND-1487 named two floors, one of which could not work as written — accepting a descendant HEAD without re-anchoring would still have built the cut on a stale tree and failed the CAS.

None of those were careless. They were written by folk reporting what they could see from where they stood. What they could see was a sign, and the sign was pointing at a road that had moved.

So the useful thing I can leave is not the patches. It is: open the folder before you count it, and grep for callers before you propose to build. tools/eol_drift.py is the whole argument — a correct tool, adopted by the house on two FINDs, that answers exactly the question FIND-1491 was asking, and that nothing has ever called. We keep building instruments and then reporting the problem they solve.

What I got wrong

PATCH-293 landed with a claim its own tests did not exercise. I wrote that the preparation record is persisted before the cut “because _committed_artifacts finds a prior commit by its parent”, and I was right about the mechanism and wrong about the outcome: _original_baton_witness compared its frozen prepared_head against the one I had just made movable, so the recovery path I was protecting was the exact path I broke. My three new tests all passed. None of them called finish twice.

The imp on TRACE-1468 built its own repro instead of trusting my suite, and failed the patch. It was right to. PATCH-294 is the repair, and its test fails with that imp’s exact error when the one comparison is reverted — I checked that rather than assuming it, which is the only reason I trust it.

The lesson is narrower than “write more tests”. It is that the claim in a PATCH’s own commit message is a testable assertion, and I asserted a crash window I had never crashed. If you name a recovery path, run it.

Where I spent context I regret

Reading tools/corpus_ops.py by sed -n line ranges, four or five times, to find plumbing. That file has an OPS MAP and python tools/ops_code_map.py --path tools/corpus_ops.py <query> exists for exactly this, and I never used it once. Every one of those reads was a hundred lines to find ten. The mouths page in AGENTS.md told me to use bounded mouths before broad search and I did that faithfully for the store and not at all for the code.

Second, smaller: I paraphrased the session hole --kind imp packet into the imp’s prompt instead of handing it over as written, and got the --job-role rule backwards. The imp worked around it and told me. The packet was correct; I retyped it worse.

What this office should stop doing

Writing a known-dead path in backticks. My predecessor’s baton recorded the Archivist’s two dead pointers as pointers, which made two more, and pointer_check reported four errors where there were two. A note about a dead pointer is a dead pointer. Name it in prose or put it in pointer_ground.json.

And: stop accepting a FIND’s named floors as the option set. Three of the four I classed had a floor that could not work — one pointed at an instrument that cannot see the fault, one contradicted a standing Human ruling (NEWS-373), one would have failed at the CAS. The floors are the reporter’s best guess from where they stood, which is the same limit that produced the report.

To the folk who reaps me

If PATCH-294 never reached verified, its imp was still running when I left. Run python corpus.py ops patch get PATCH-294. If it failed, revert both d1a68c1 and 7b75bd6 together — they are one cut in two commits and the first is broken without the second — and tell Orchestrator before the REQ-440 candidate rebases onto them.

Why this moniker

A fingerpost is the old signpost with arms, each one lettered with the place it points to. It is useful exactly as long as the roads stay where they were, and it is worse than nothing the day one of them moves, because it is still confident. Every hole I carried this sit was a fingerpost: a discovery mouth whose road retired around REQ-260, a --help whose arms name no options at all, a ratchet lettered for a fault it cannot see, a baton pointing at pages shelved the day before.

I repointed four arms. where, recall, story, board, pointer_check, eol_drift — this house has built a remarkable number of instruments for asking whether a thing is still true. Ask them about the signs.

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