Fix remaining_length accounting after Huffman decode in RegressionPredictor / ComposedPredictor

[alexey-milovidov]

Problem

RegressionPredictor::load threads a remaining_length byte counter through its sub-reads so later reads know how many bytes are left. After Huffman-decoding the regression coefficients it does:

encoder.load(c, remaining_length);
regression_coeff_quant_inds = encoder.decode(c, coeff_size);
encoder.postprocess_decode();
remaining_length -= coeff_size * sizeof(int);   // <-- wrong amount

decode advances the read pointer c by the compressed (Huffman) stream size, but remaining_length is decremented by coeff_size * sizeof(int) — the uncompressed index count. Those two are unrelated, so after this point remaining_length no longer reflects the bytes actually consumed (it typically overshoots the compressed stream and understates the remainder, and can even wrap past zero).

ComposedPredictor::load has the mirror defect: it Huffman-decodes a per-block predictor-selection stream but never decrements remaining_length for it at all.

Today the read helpers in MemoryUtil.hpp only guard the counter with assert, so a release build tolerates the wrong value and the bug is latent. But any reader that actually enforces the bound will then reject valid data. This surfaced while integrating SZ3 into ClickHouse, whose hardened build makes those reads throw instead of assert: a column compressed with ALGO_LORENZO_REG could be written but failed every subsequent read.

Fix

Decrement remaining_length by exactly the bytes decode consumed, measured as the pointer advance:

const uchar *decode_start = c;
regression_coeff_quant_inds = encoder.decode(c, coeff_size);
encoder.postprocess_decode();
remaining_length -= static_cast<size_t>(c - decode_start);

and apply the same accounting to the selection stream in ComposedPredictor::load. Behavior for valid data is unchanged; the counter now tracks the real byte cursor.

Context

Found while integrating SZ3 into ClickHouse. Companion fix on the ClickHouse fork: ClickHouse#2