feat(evals): close coverage gaps — 6 authored objective sets + StrategyQA + run batches

Judge-free objective binary sets (balanced ~15/15, ~30 each) for skills that were
previously unmeasured or proxy-only, consumable by run-routing-data.js:
- reversibility (one-way vs two-way door)
- first-principles (fundamental limit vs convention)
- theory-of-constraints (is the proposed stage the binding bottleneck?)
- second-order (does the action achieve its goal once 2nd-order effects resolve?)
- map-territory (does observed behavior contradict the stated docs/spec?)
- margin-of-safety (is the provisioning buffer adequate for the stated risk?)

Also: point the strategyqa ingest source at ChilleD/StrategyQA (the prior
voidful/ ref was unserved) with balanceScan — a real objective multi-hop set for
second-order. Add run-confirm.sh (red-team/fermi at power) and run-rerun.sh
(the 7 new objective evals). HF datasets gitignored per policy.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>