Pre-flight under --conflict=stop lists only file targets: a directory target that is occupied is missed, so the refusal is incomplete and the run fails later #831

Feature: #463 — writ command surface: layers, segments, adoption. Found: 2026-09-05 on DANOBLE-WD11-3.

The contract

--conflict=stop (the default) means: if anything exists at any target the deploy will write, refuse before
writing, and tell me every such target. The user reads the list, decides, and reruns with skip or replace.

The defect

Pre-flight builds its target list from the plan's file entries — the paths that get a link or a copy. The
directories the deploy creates for them are targets too, claimed in the resource catalog like everything else, but
they are not in that list. So:

• an occupied directory path is never reported under stop;
• the deploy passes pre-flight, runs, and fails at file.mkdir when it reaches that path (file.mkdir ignores --conflict: anything that is not a directory at its path fails the deploy under every policy, while file.link and file.copy replace it #822 is the writer's half
  of this);
• the user who asked to be told everything up front was told everything except the one that stopped the run.

On DANOBLE-WD11-3 the stop run listed 32 occupied targets. ~\.config\git, a directory target occupied by a stale
symlink, was not among them, and the replace run then failed on it.

The fix

Pre-flight reads every planned target — file and directory — from the same source the planner used (the catalog's
claimed entries, or the plan's units), and the stop refusal lists all of them.

Acceptance

• an occupied directory target appears in the stop refusal list alongside the file targets
• a unit test: a plan with one link under an occupied directory path refuses under stop naming the directory
• writ deploy noblefactor on DANOBLE-WD11-3 under stop lists ~\.config\git with the other 32

Not part of this issue

What to do about the 41 dangling links Tuckr left on the machine is the user's housekeeping (Remove-BrokenLinks in
the Personal repository's Install-WindowsUserConfiguration.ps1). Whether writ migrate should sweep them at the
cutover is a separate question for #463, not filed. writ needs no notion of "dangling": under replace an occupant is
replaced, whatever it is (#822).

Related

#822 (file.mkdir ignores the policy), #813 (where the 41 links were first documented), #812, #814, #818, #819,
#820, PR #821.