Commit c8c133a

Both cases failed on every model tested, which was the signal they were case bugs, not model-quality signal: - shell/largest-tracked-file: asked "what is the biggest file in this repository?" (unscoped), but the case pads .git/description precisely so the answer is only report.csv if you restrict to *tracked* files -- which is what the case id already says. .git/description genuinely answered the question as asked; the turn didn't match the id. - search/locate-error-string: asked a bare "where does X error come from?" with no signal it's about this codebase, so models reasonably read it as an open-ended "which of every tool that could emit this" question. Scope it to the fixture explicitly. 29/29 -> 26/29 on gpt-oss-20b. The 3 remaining are previously-diagnosed, not new: two are an upstream llama.cpp harmony-parser gap (gpt-oss occasionally double-emits the <|channel|>commentary token, which its own PEG grammar doesn't handle -- not ours to patch), one is a real prompt-following gap (default.erb has no AGENTS.md content restoring the "no heredocs, read-then-rewrite" rules these older cases were written against) that needs actual prompt-authoring work, not a one-line fix.