Builds balance suite against the Power sheet

Phase 10 of #147/#148 (docs/builds.md § 15, § 20.4). uo-loadtest builds: duels and pack fights between saved loadouts (the ±10 % damage roll is the only randomness), damage share per bucket and source, tier curves compared with the workbook's Power sheet; the targets of § 15.4 (complete T1 power 3–8; a complete T5 build beats 9–11 complete T1 builds of its kind all in and 6.5–9 on gear alone, #149) and the mirror-fight drift; CI keeps the catalogue's example builds within their targets; a playtest wave and a tuning pass.

Done when the measured power curves match the Power sheet within ±20 % and the tuning is recorded in catalog.py.

Also (user directive, 2026-10-29): a full top-tier party fight must last long enough to read on its own numbers, so the Thornford showcase's HITS_SCALE (×3 hits, uo_combat::pvp) goes away; count the effective builds (sample thousands of plausible builds, duel, pack and party fights, count the ones on the Pareto front) as a diversity metric; the elite test hero's picks (#171) sit near its thresholds.