Name the failure when task setup dies before the worker spawns

[mlava]

Independent of #35/#36 (different code region); trivially mergeable in any order — the only shared file is ringer.py in non-overlapping functions, plus a one-line adjacency in print_summary.

Problem

Failed tasks keep their worktrees (by design, for post-mortems). A follow-up run with the same run_name then collides with the stale taskdir and dies pre-spawn: verdict ERROR, elapsed_s: 0.0, empty log_tail, empty check output — and no message anywhere. In the field (2026-07-14, protocol-spec-fixes-20260714T131326Z) it took a full diagnosis cycle (git worktree list) to find what one sentence could have said. The prepare-error string existed internally, but only landed in a model-log row's notes — not the worker log, not the task record, not the summary.

Fix — three surfaces

1. The collision message names the exact unblocking command: git worktree remove --force <taskdir> (instead of a bare "already exists").
2. _record_prepare_error stores the reason as a new setup_error field on the task record (included in the run-state payload, so the HUD and post-mortems see it) and appends it to the worker log — where log_tail, activity, and humans look first.
3. The run summary prints a setup failures (no worker was spawned) section listing each affected task with its reason.

Test

End-to-end worktrees run against a temp git repo with a pre-created stale taskdir (tests/test_setup_error_diagnostics.py): asserts the ERROR verdict now travels with the collision message in the summary, the worker log, and the run-state record's setup_error. Full suite: no new failures (the pre-existing test_design_reference/test_scoreboard_page failures reproduce on clean main).

Related follow-up not in this PR: an explicit --reclaim flag to auto-remove stale worktrees from prior runs of the same manifest.

🤖 Generated with Claude Code

[justfinethanku]

Merging this one — thank you. The instant-fail-with-empty-everything signature has personally cost us a diagnosis cycle (it's been folklore in our MODEL-NOTES since), and routing the reason to all three surfaces an operator actually checks — worker log, run state, summary — is the right shape. It also cleanly separates never-spawned failures from model evidence, which matters to how we read the scoreboard.

One maintainer-side adjustment before merge (your authorship stays on the commit): the recovery command is now printed only when the stale path is a registered worktree (.git file present), repo-qualified (git -C <repo> ...) and shell-quoted, so it's paste-safe from any directory. A plain directory in the way — where git worktree remove would itself fail — now gets named as exactly that with no command we haven't verified. Your test was restructured to cover both branches, with the worktree case using a real registered stale worktree.

Full suite green at 215 tests. Also: the joint-design invitation on #33/#36 is genuine — this PR's _record_prepare_error surface is adjacent to that work, so whoever lands the unified breaker will rebase over this small diff easily. Thanks for the steady quality.