Targeted search with iterated local search; future-window lower bounds (content of #43, #44) - #46

#43 and #44 were merged into their stacked base branches (exact-certificates, targeted-search-speed) after those had already merged, so their changes never reached main. This PR cherry-picks the two commits onto current main without changing them.

Summary

• Targeted search (Targeted line-split search: iterated local search; faster kernel LP #43): targeted_line_split_seeds runs iterated local search (2 chains x 30 hops) after COBYLA and polishes only the 3 lowest-entropy decompositions. _kernel_lp_beliefs is vectorized (same LP, faster to build). On the widest survey bracket it reaches the 0.970-bit generator in 8 of 8 seeds. Survey v10: ((0,1),(0,2),(0,1)) upper 1.037 -> 0.970 and ((1,2),(0,2),(0,1)) upper 1.316 -> 1.091.
• Window bounds (Future-window generator problems: C_g lower bounds without a reverse machine #44): window_spec(forward, L) and _GeneratorProblem.from_spec. Generator equations with the reverse causal state replaced by the next L symbols, giving lower bounds on C_g from the forward epsilon-machine alone.

Test plan

• The generator, certificate, window-spec, state-label and bidirectional suites pass on the combined code (160 tests).
• ruff check / format; ty (no errors).
• Code matches window-spec except for what main already gained from State labels beyond 26: AA–ZZ, AAA–ZZZ, ... #45.

Made with Cursor