feat(m5): powered primary batch — 7 gated skills evaluated, all NO-LIFT

Gate-checked and powered 7 approved skills (red-team, occams-razor, five-whys-plus,
systems, archetypes, map-territory, kepner-tregoe) using solver claude-sonnet-4-6,
CONC=4, isolation ON (EVAL_RUN=m5-primary). All are post-edit runs against M4-reworked
skill versions on frozen calibrated splits.

Results (all NO-LIFT, none passed ≥5pp + p<0.05 primary gate):
- red-team: +1.4pp p=1.0 (security-adversarial decisive, n=70)
- five-whys-plus: +0pp p=0.752 (SWE-bench, n=150)
- systems: +1.3pp p=0.724 (SWE-bench, n=150)
- occams-razor: +4.7pp p=0.096 (SWE-bench, n=150, directional)
- map-territory: +2.7pp p=0.221 (exploratory: surface-mismatch)
- kepner-tregoe: +2.7pp p=0.289 (exploratory: surface-mismatch)
- archetypes: +0.9pp p=1.0 (systems pairwise, n=117, quarantine candidate)

map-territory and kepner-tregoe tagged 'exploratory: surface-mismatch (powered on
fault-localization, not native surface)' — their no-lift is NOT an honest kill.
archetypes was an in-band quarantine candidate given a fair calibrated chance.

Scorecard (JSON+MD) updated with full evidence tuples + provenance + verdict taxonomy.
New run-binary-decision.js runner for systems-product-strategy-pairwise family.

Assertions fulfilled: VAL-POWERED-001/004/005/006/009/010/012/013,
VAL-DATASET-001/009, VAL-SKILLS-008.