notes: the twenty-third pass — the desktop renamed out from under this file, and a mirror one repo over

[JulienMartel]

What

The twenty-third pass on notes/options-roadmap.md, plus two stale rows the
rename left behind in AGENTS.md and notes/hausfold-rename.md.

The audit half. haus#364
(decision 10) merged four hours after the twenty-second pass's own PR and dropped
the name nebelhaus: the desktop is hacker, everything else in the layer
is haus. Nothing this file tracks as work moved — the rename configures
nothing differently and its three compatibility seams mean no consumer has to
act. What moved is every sentence that says the word, so the naming banner is
amended and the body stays un-rewritten, per that banner's own rule.

Three findings, one of which shipped code.

1. A translation rule for a historical document is only as good as the token it
   keys on. The 2026-08-08 banner said "read every nebelhaus.* as haus.*" —
   keyed on the dot, because what was renamed then was an option namespace.
   Decision 10 renamed the bare word. Over the body proper: 137 hits, the
   dotted rule translates 63 of them correctly and 2 of them wrongly
   (nebelhaus.com and a filename, both of which stay). The workshop's own
   AGENTS.md had already written that trap down — for the layer, one repo over,
   in the same week.
2. "Audit against the repos, not against memory" has a third thing in it: the
   local checkout is memory wearing the repo's clothes. ~/code/workshop/haus
   still held desktops/nebelhaus.nix 46 minutes after notes: §5.2 closes, and the route that outlived its config #364 merged, so the disk
   would have answered every question with yesterday, in the confident voice of a
   file that exists. bench status fetches before comparing @{u} (bench:547)
   and would have said ↓ 1. Two rows for §5.14's shapes table, since the second
   half of that finding was a negative claim ("bench never fetches") produced
   by a grep whose pattern couldn't match git -C "$dir" fetch.
3. The lock ripple carries the binary and nothing carries the grammar it
   speaks. haus.pounce.items validates keys against a hand copy of pounce's
   ItemTarget; pounce added shortcut:<uuid> (pounce#80) and haus's lock moved
   to that exact merge commit 79 seconds later — so the layer has been refusing a
   key its own daemon accepts. Fixed, with a seventh check, in
   haus#365. §5.9(b) amended: it
   asserted that validation as a feature.

Why the two AGENTS.md / rename-plan edits are here

Both are shape 2 in this file's own table (a closed claim, later falsified) and
both were found by the assurance read: AGENTS.md and hausfold-rename.md
§2.2's do-not-sweep table still call the four state dirs "deliberately held",
left behind by the very commit (workshop#372) that amended the row four lines
above. §11.3 moved all four and left symlinks. Fixing them where they were found
beats filing them, which is this file's own rule about follow-ups without owners.

Verify

Prose only — nothing here evaluates. What was checked, and by whom:

• Two clean-context assurance subagents, one per repo, each given only the diff
  and that repo's AGENTS.md. Eleven findings between them, four ≥3/5, all
  folded in — including both count claims in the first draft of the status block
  (a pair of counts taken over two different regions of the file), a check that
  could go green while the validator it protects stayed broken, and haus's CI
  census.
• Every quoted line re-read at its source: the builtinModes comment,
  AGENTS.md's bare-word warning, bench:547, pounce's ItemTarget.
• Counts re-derived rather than quoted, over one stated scope, with the command
  that produced them.

Watch out

• haus#365 is open, not merged. The defect finding 3 describes in the past
  tense is still live on main until it lands; the notes say so and name the PR.
• The 137 / 63 counts are scoped to the body (§1 onward) on purpose, so a
  pass that writes the word fifteen times doesn't inflate its own number. A
  whole-file count is a different, larger, self-referential figure.
• §5.14's shapes table is ten rows now. The two new ones are last, and the
  paragraph below the table that talks about "the sixth shape" is unaffected —
  it names the row, not the index.

🤖 Generated with Claude Code

https://claude.ai/code/session_018g5zLnGtTtVDkLjfJ7rNCX