D122 — PCBWorld D3 run as PCBench boards with Freerouting rerun

Decision

PCBWorld-Bench's real boards (D3) are PCBench boards, so A1 became: run the 119 published test boards (99 D3-A, 10 D3-B, 10 D3-C) and re-run Freerouting 2.1.0 with the paper's settings under this repository's judge as the only like-for-like baseline; the paper's other rows are quoted, not compared. TraceMaker: 97/99, 10/10, 9/10 clean. --txn 1 helps on non-breakout boards (+118 connections on 33 failing tier B-D boards, eight variants, equal work), --relax is a rescue (two boards completed, 509 connections fewer in all); neither changes a default yet. --relax, --rounds, --round-work and --ramp-from now run a portfolio of one with --variants 1

Alternatives

Reproducing the paper's board conversion (needs a KiCad 9 pcbnew built from their engine); a "better than" claim against KiCadRoutingTools, OrthoRoute, the agents and the RL policies (different judge and rule files, no per-board results); making --txn 1 the default now (it changes the default portfolio and needs tiers A to D at equal wall time first)

Reason

Doc 18 assumed D3 was an independent set; reading the paper showed it is a PCBench subset with relaxed rules. The paper's Freerouting 2.1.0 scores 0.75 (one run) here at 0.670, so the judges differ and only same-judge numbers mean anything; the experiment that found --relax ignored with one variant is the reason for the second change