restore winning v8 config on main; README: final-day results (v11-v13 all negative)

v11: beam search (num_beams=4) on the answer pass; single change on the 0.2245 v8 config

restore batch=1 default (the winning configuration); v10 batch=4 variant captured at commit 1aa6a60c

v10: winning config but batch=4 (tests empirically whether left-padded batching costs exact match; batch=1 winner is commit f283a4a3/8aa238da)

v9: winning baseline config kept as default; document the 0.2245 rank-1 result; make baseline batch size configurable (default 1, the winning setting)

v8: faithful replication of the organizers' baseline (batch=1, no padding, 512 tok, naive parse, no N-forcing) + repetition_penalty=1.0

v6: pass repetition_penalty=1.0 explicitly (model ships 1.05, silently applied under greedy)

v5: revert to the best-measured config (submission 2) + match_letters assignment solver

v4: mirror the problem's own gloss style (36/36 verified) instead of forcing capitalisation; equation-aware text_to_num; digit+letter-case normalisation (99.9% safe on 920 golds)

v4 981a->mirror: answer style copies the problem's own gloss style (measured 36/36); equation-aware text_to_num; one-to-one match_letters

answer-style guidance per task type (closes the EM/chrF gap); relax pass margin to 1.1

fix: greedy-anchored voting (symmetric vote cost 4x EM on mock), truncation diagnostic

script.py: aligned answer blocks, self-consistency voting, task-type format hints