Add run --baseline: execute every check against the unmodified tree

[mlava]

Independent of #35/#36/#37 — new function + one flag + one dispatch branch; no overlap with those diffs.

Problem

The check is the product — but nothing exercises the checks themselves before workers spawn. In a real 12-lane fix swarm (2026-07-14), two lanes failed on verify assertions that were wrong about the pre-change tree: one demanded a status its own lane spec forbade, and one was unsatisfiable on a pristine repo (a tsx -e CJS/ESM quirk). An honest worker burned ~100k tokens against the unsatisfiable one before anyone knew the check was the bug.

Feature

ringer.py run manifest.json --baseline executes every task's check against the unmodified tree and reports — spawning no workers:

Baseline: executing 12 check(s) with no workers spawned.
unchanged-behavior       baseline: pass (rc=0)
new-behavior             baseline: FAIL (rc=1)
    FAIL: built-by-worker.txt not present

baseline: 1 pass, 1 fail, 0 error of 2 check(s).
Reading the results: a FAIL is EXPECTED for assertions that demand the
NEW behavior workers will build. A FAIL on an assertion about UNCHANGED
behavior means the check itself is broken and will burn worker attempts
against something no model can satisfy — fix the check before spawning.

Design notes:

• Spawns nothing — dispatches before preflight_engine_bins on purpose, so a missing engine binary can't block baseline
• Fresh scratch taskdir per task (a detached worktree when the manifest uses worktrees), removed afterwards; nothing leaks into the manifest workdir or the repo
• No model-log rows
• Checks run for real, including any exports they perform — same semantics as at run time (reuses Verifier, including expect_files and CHECK_TIMEOUT_S)
• Exit 0 regardless of failures: baseline reports; deciding which failures are expected is the orchestrator's judgment, and the report says exactly how to read it

Test

End-to-end (tests/test_baseline_mode.py): worktrees manifest on a temp git repo, a deliberately nonexistent engine binary, and two checks — one asserting current tree state (must pass), one demanding a worker-built file (expected FAIL). Asserts the report lines and excerpt, the guidance text, exit 0, an absent model log, and zero leaked taskdirs or worktrees. Full suite: no new failures (the pre-existing test_design_reference/test_scoreboard_page failures reproduce on clean main).

🤖 Generated with Claude Code

[justfinethanku]

Merging — thank you. This is the most doctrine-aligned PR in the queue: "a check that cannot fail is trusting the worker with extra steps" has been in the README for weeks, and --baseline makes it a one-command preflight instead of a lesson you learn at ~100k tokens a lane. The field story in your test docstring is exactly why we wanted it, and reusing the real Verifier so baseline semantics match run-time semantics was the right call.

Maintainer-side before merge (authorship stays yours): resolved the parser conflict with current main (--baseline now sits alongside --allow-noncanonical-route), added the same task-key containment rule the real run path enforces so a hostile key can't escape the baseline scratch root (with a regression test), and made failed worktree removals print per task plus a summary warning with the exact git worktree prune cleanup — a clean-looking report must not hide leaked state. Full suite green at 217 tests.

Also queued on our side, sparked partly by this: a structured failure-class field on eval rows so "the check was broken" vs "the model failed" vs "the engine was down" becomes sliceable scoreboard signal. Your #35/#36 verdicts feed the same taxonomy — the joint-design thread on #33/#36 is the place if you want in on that. Thanks again for the consistently excellent work.