fix: four papercuts in operating the harness (#147, #148, #149, #153)

[thedancingdeveloper]

None of these is about model quality. All four are about the harness being
usable by someone who has not read its source, and all four were hit within
the first hour of pointing it at an unfamiliar repository (#56).

Stacked on #151 because it touches the same files.

#147 — a run destroyed the checkout it was given

Headless works in place and begins each attempt with git checkout -- . and
git clean -fd, so a modified tracked file was reverted and an untracked one
deleted, neither recoverably, and no document mentioned it. A real import came
within one command of losing 136 uncommitted files including two whole crates.

Preflight now refuses a dirty checkout before anything is claimed, doctor
reports it before you get that far, and --allow-dirty overrides it — loudly,
with the detail kept in the report, because the whole point is that the loss is
silent and irreversible. Session mode is unaffected: it has its own worktree.

#153 — the role flags were a seed pretending to be a setting

After any run had stored a map, --implementer was inert; two runs were spent
on a model that had been explicitly replaced on the command line. The warning
existed and went to stdout, which is block-buffered into a pipe — so it
arrived out of order when it arrived, and vanished entirely when the process
was killed.

It is a log line now, it names both models, and --reroute makes the command
line win. The second symptom is fixed too: a complete stored map no longer
refuses to start for want of flags it did not need, which had made the flags
simultaneously required and ignored.

#148 — plan --dry-run demanded a GitHub repository

To answer a purely local question. --repo is now optional under --dry-run,
which prints the items it parsed and stops.

#149 — verify: accepted a test filter that matched nothing

cargo test -p gateway secure_cookies ran zero tests, exited 0, and adoption
reported the item as already delivered. The harness cannot tell that apart
without reading another ecosystem's output, and should not try — so this is
documented, with the failure shown, next to the sentence that promised exit 0
was evidence.

Docs

USAGE.md gains what headless does to your checkout, what the role flags
actually are, and how to write a verify: that fails when the work is absent.

All four gates green.

🤖 Generated with Claude Code