Speed shared measurement cell assembly and CSV export with stable-ABI primitives #239

Current wide-producer spreadsheet export for the representative one-well ImagingFlowCytometry pipeline spends about 0.94 s assembling tables and 0.79 s serializing CSV, ~2.04 s overall. This cost repeats with wells and selected cells.

Saved current producer input: 14 MB, 31 measurement payloads, selected output 1,800 rows / 11,550,064 CSV bytes, SHA256 3e00436ae1500047fd2605021ae055492b17dbe0b4aa4ddbe2f62af6f8e073be. This is current wide-form data; the old eight-well long-form artifact is not the current frontier.

A shared compiled cell-assignment / CSV prototype preserves all nine saved-input observations byte for byte. Alternating replays: current Python ~1.93 s, compiled assembly alone ~1.80 s, compiled assembly plus CSV ~1.445 s. The bounded improvement is ~25% export time, not an established whole-pipeline speedup and not a fix for first-ever cold compilation. The existing accumulator must retain row identity, scope and qualifier ownership, with one shared conflict-assignment implementation and existing scalar-equality authority. Keep Python numeric string rendering and CSV quoting, nonfinite handling, source encodings and field order exact; do not add an import fallback or duplicate a backend roster.

Production admission requires meaningful adversarial equality/sparsity/projection-order and CSV parity tests, stable-ABI wheel consumption, and ordinary fresh endpoint benchmark with exact output checks. Current release-floor test also assumes setup() has only cmdclass, although main already builds the native granularity extension through ext_modules; that stale guard must admit build-only extensions while continuing to reject dependency/metadata duplication.

Related: #162 remains open for kernel/timing scope. Prior #194 and #216/#219 addressed different long-form/wide-producer bottlenecks and are already merged.