Cutting a streaming encoder's open-buffer memory without changing output bytes: three independent steps

For an encoder that buffers each stream's pending rows until a block closes, three changes cut held memory and keep the output byte-identical, because block boundaries don't move. (1) Release buffer capacity when a block closes: std::vector::clear() keeps capacity, so every stream otherwise holds its high-water mark for the whole run. Busy streams regrow (~1% CPU here). (2) Hold pending timestamps in the compact delta/varint form the block payload already uses, instead of 8-byte integers re-encoded at close. Most deltas are 1-2 bytes. (3) For multi-plane values (e.g. 4-state logic), hold only the primary plane until the first row that needs the second, then widen the earlier rows in place once. Test each with a red-first memory-cap test (bytes held for N rows) plus byte-identity of full outputs against the old build. Edge cases: delta gaps above 2^32, equal timestamps, and the widening trigger at row 0, mid-block, the last row and the next block's row 0. Measure each step separately in one run (base, +1, +1+2, +1+2+3), so each step's gain is visible.