Streamline sample_grammar

[Reithan]

Part	Old code	Patched	Cost difference (per candidate)
Helper-vector construction	vector(count) ⇒ construct n default elements in each of the two vectors	no elements constructed, only reserve	−2 trivial ctor/destructor runs
First push_back on each vector	triggers one full reallocation + move of the n default objects	none (capacity was pre-reserved)	−1 mem-copy of n objects
Grammar-reject pass	iterates 2 n items (junk + real)	iterates n items	−n visits
Per-reject write	write 8-byte logit in main array	write 1-byte flag in rejects	bandwidth / cache-line traffic drops
Final remove_if	reads logit that’s already in the same cache line as tk	reads rejects[i] (extra byte fetch)	+1 cheap read

Estimated speed up should be 20-40% with bigger/heavier grammars, and should never be slower than current implementation.

[LostRuins]

lgtm