Ticket graphs

Amends #200: NO module rewrite is single-shot. Each rewrite ticket produces AT LEAST 5 INDEPENDENT implementations, diverse by construction (different models — gemma3:27b / qwen3-coder / grok / claude — and/or different approaches). For each of the 5+: (1) it MUST pass behavioral equivalence (#200 exemplars) — a GATE, not a score; (2) PERFORMANCE is CAPTURED — runtime, memory, allocation, hot-path counts — via the born-instrumented telemetry (#203), on a shared benchmark harness, real numbers not estimates. Then an ALL-AGENT ANALYSIS (a judge PANEL — multiple independent agents/models, each scoring blind) rates every candidate on the RUBRIC: - CONFIGURATION OPTIONS: how completely behavior is driven by the control-config object (nothing hardcoded); - EFFICIENCY: the captured performance results (objective, from the benchmark); - HUMAN READABILITY: clarity, naming, structure, cognitive load (scored by readability-auditor agents). Aggregate the panel scores (median across judges, disagreement flagged), rank the 5, SELECT THE WINNER, and GRAFT the best ideas from runners-up. A per-module DEEP-DIVE record (docs/rebuild/<module>.md) captures: the 5 candidates, their perf tables, the panel scorecards, the winner + rationale. This is the tournament pattern the prototype already gestured at (code_wrapper make_candidates + critic, #185). The WINNER is what the rewrite ticket settles with.

— browse congruency source (forge-parked tree)

— #207 Build: competitive-rewrite tournament harness — generate 5+ candidates, benchmark, all-agent score, select winner, record deep-dive (spawned · DONE)