[codethinki]

Replaces the hand written timings with a Google Benchmark target, lbs_solver_bench.

puzzles

• assets/nyt_puzzles.txt: the 558 nyt puzzles from 2025-03-25 to 2026-10-05, one <date> <letters> per line
• taken from the daily archive in cbrady2718/letterBoxedSolver (nyt_data.json), sides only
• split into solvable and unsolvable with our word list (556 / 2)
• lbs and gpt have to agree on solvability for every puzzle, otherwise the benchmark exits with 1

benchmarks

• every puzzle is its own benchmark, so warmup, iteration count and DoNotOptimize are handled by Google Benchmark
• lbs/load, lbs/split: the word list once
• lbs/filter_and_hash, lbs/solve, lbs/total: per puzzle, word list already in memory (the #embed case)
• lbs/total_with_load, gpt/total_with_load: per puzzle including reading the word list
• a custom reporter prints min, p50, p90, p99, max and mean per benchmark, --benchmark_out has the per puzzle data
• --benchmark_filter=2026- etc. selects puzzles by date
• default min time is 0.01s per puzzle, the full run takes about 50s

other

• gpt_runner.hpp moved into bench and reads the file with std::ifstream instead of a windows file mapping
• benchmark>=1.9.5, the 1.9.0 port at our baseline passes -Werror=old-style-cast to cmake which cmake 4 rejects
• runner and solver.real_world no longer time anything, real_world checks that a solution is found

results (solvable puzzles, p50 / p99 in us, Core Ultra 9 285H)

	gcc 16	clang
lbs/split	1127	473
lbs/total	290 / 504	298 / 664
lbs/total_with_load	1357 / 1642	822 / 1174
gpt/total_with_load	962 / 1110	983 / 1458