run_scenario always ran `bd init -p <prefix> --quiet` before a scenario's
steps. 43 deterministic catalog scenarios start with their own
`init --prefix X --non-interactive ...`; in 40 of them that step always hit
"This workspace is already initialized" (exit 1), pinning the double-init
refusal instead of a real init. The other 3 fail flag validation first,
which is what they test, and are unchanged. Their later steps (the commands
they are named for) ran against the harness's own pre-init and are
unaffected.
A scenario whose first step's command token is `init` now owns
initialization (Scenario::owns_init) and starts in an empty workspace; every
other scenario keeps the pre-init, and a later `init` step stays a re-init.
The pre-init decision is the pure harness_init() helper, and the step loop
is run_steps(), which takes the command runner as a closure.
normalize() now collapses the `Repository ID`/`Clone ID` values a real init
prints (hashed from the random workspace path) to <REPO_ID>/<CLONE_ID>.
init_proxied_server_not_implemented is marked deterministic:false: its
premise is stale and, un-masked, it starts a managed dolt sql-server.
New tests cover the owns_init rule, the harness_init decision, the
run_steps argv sequence (via a recording closure) and the fingerprint
normalization. None spawns a process, so none can hit the write-then-exec
ETXTBSY race against the suite's process-spawning tests.
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>