hq-pack-bake-off

/bake-off — head-to-head model provider tests. Races two or more coding agents (Claude vs Codex vs Grok vs Gemini vs any headless CLI it finds on your machine) against the IDENTICAL build spec on isolated git branches/worktrees. Claude's own entry is built by a fresh subagent that sees only the spec — same information diet as the externals. Every hard requirement is verified by running a mechanical check (command/URL/grep), never by the orchestrator's opinion. Entries are presented with randomized, sealed blind labels 🅰/🅱/🅲 plus objective columns (verify pass/fail, diff stats, wall-clock, cost where reported); the human judges, then the mapping is revealed. Verdicts accumulate in a running ledger — `/bake-off stats` tallies wins, DNFs, speed, and cost per provider over time.