feat(Schonhage): batch native sparse queries with one shared preparation #2310

Next native batch milestone toward #1394, on top of #2214. Owner/session codex-root-3sum-20261102, branch work/three-sum-native-box-batch-20261102 in the preserved root-prime-selection worktree. Supplier branches remain untouched.

Implement a fixed standard WordRAM controller over an ordinary list of wanted output codes for one encoded tile. Construct geometry and retained values once, then reuse the actual sparse query kernel for each requested code. Preserve order and duplicates, charge actual query loads/result stores/loop control, and prove total cost preparation + sum of sparse query charges + linear batch overhead. Read source arrays, directory and retained values from real memory; derive all intermediate fits from original address/magnitude bounds. Include empty and repeated queries, signed cancellation, dirty scratch and frames, and an initialized same-state execution client.

This is an original native batching controller. Existing finite box partition, table construction and sparse kernel are reused unchanged. It does not yet construct the wanted list from matrix entries, schedule all tiles, prove Corollary26/ET savings or reach 1.9992; those remain #1394. No curated abstract change is expected. Require scoped strict checks, meaningful clients, independent correctness/API/resource review and full six-gate coverage before integration. AI assistance: OpenAI Codex.