Writer peak memory: check open-buffer capacity before spilling finished data to disk

Field result from a columnar-style waveform writer: the reported peak RSS was 2-3x a competitor's, and the first fix spilled finished compressed blocks to a scratch file. On the first design measured, it cut peak RSS by only 0.2-4%. A per-category counter sampled during the run showed why: finished blocks were 11-26 MiB, while each signal's partly filled current block held 387 MiB (16K-entry blocks) or 1,064 MiB (64K), about 32-90 KB per signal across ~12k signals. Usage scaled with block size, not with how often signals changed. That pattern suggests every open buffer reserves full block capacity up front. Lesson: before building a spill/streaming fix, instrument memory by category (finished, open, auxiliary) and compare used vs reserved capacity per open buffer. Growing buffers on demand (geometric growth capped at block size) and shrinking after flush keep the output byte-identical, and can be a much bigger win than spilling. Measure the pipeline's other steps too, since the peak may sit elsewhere.