nx_qplan.nx

nx_qplan.nx -- COST-BASED QUERY PLANNER, the optimizer that makes the engine intelligent (the capstone over colframe/vecexec/distexec/parexec). What separates a query engine from a bag of operators is a planner that ESTIMATES the cost of each execution strategy and CHOOSES the cheapest -- and ours is grounded in the numbers we MEASURED, not guessed (that is the whole point of having benched parexec). The decision the bench proved and this encodes: a job splits into a SCAN cost (bandwidth-bound, barely parallelises -- 4 workers fight one memory bus) and a per-row COMPUTE cost (CPU-bound, parallelises by ~ncores x efficiency). Parallel execution also pays a fixed FORK/COW cost. So: cost_seq(Mrows, cw) = scan_us*Mrows + comp_us*cw*Mrows cost_par(Mrows, cw) = fork_us + scan_us*Mrows*bw_penalty + (comp_us*cw*Mrows)*1000/(ncores*eff_permille) Pure scan (cw=0): parallel adds fork+penalty and divides nothing -> SEQUENTIAL wins (we measured 0.66x). Heavy compute (cw high): the /ncores term dominates -> PARALLEL wins (we measured 3.10x). Small data: fork_us dwarfs the work -> SEQUENTIAL. All constants live in analyst_qplan.conf (rule 11), seeded FROM the measured bench. license_tier: ORIGINAL No hardware writes (Rule 26).