[ColinLeeo]

The previous compatibility fixtures cover selected codecs with 300-row or NULL-page samples. They do not exercise all implemented common codecs, large datasets, or Java's default BOOLEAN RLE encoding alongside other columns.

This expands the matrix to 47 implemented common type/encoding pairs, six common compression algorithms (UNCOMPRESSED, SNAPPY, GZIP, LZ4, ZSTD, LZMA2), and 32 / 4,096 / 1,000,000 rows. Each core fixture includes a separate boolean_value column and independent, fixed-seed 10% NULL masks for the main value and BOOLEAN columns. Dedicated legacy NULL-page cases retain their structural NULL patterns. Java uses its default BOOLEAN RLE; C++ uses its implemented BOOLEAN PLAIN writer. Both readers verify actual chunk codecs, values, NULL positions, timestamps, and column alignment.

• 846 core fixtures per writer, plus 72 existing NULL-page fixtures.
• Java additionally retains 36 explicit max_point_number=0 fixtures: 954 Java files and 918 C++ files in total.
• Compare manifest coverage sets so missing combinations cannot silently pass.
• Enumerate the implemented intersection verified against both encoder/decoder factories, excluding enum-only placeholders. Actual fixture generation and readback exercise every included combination.
• Shard CI by compression and enable SNAPPY/GZIP. Run the target Surefire method once, and use 4,096-row pages for scale cases while retaining 64-row NULL-page regressions.

BOOLEAN RLE support (#1019) has merged into develop. The newly exposed codec defects are addressed independently in #1022 (DICTIONARY embedded NUL bytes) and #1023 (SPRINTZ extreme transitions, FIRE arithmetic, and block boundaries). This matrix PR remains draft until those fixes merge.

Baseline failures (before the independent fixes)

The complete 32-row matrix reproduces existing failures across all six compressors:

• SPRINTZ INT32/INT64: extreme-value transitions decode incorrectly in both directions.
• SPRINTZ FLOAT/DOUBLE: negative zero loses its sign bit in both directions.
• DICTIONARY TEXT/STRING: C++ truncates dictionary strings containing embedded NUL bytes.

Million-row shard checks also reproduce these failures. This draft preserves the failing cases so they remain visible as merge blockers. The failing inputs are retained; fixes are provided in #1022 and #1023.

Validation

• Complete fixture generation passed: 954 Java files (295,239,744 rows) and 918 C++ files (283,190,208 rows).
• Full Java-to-C++ audit: all 954 fixtures attempted; 846 passed, 108 failed (72 SPRINTZ + 36 DICTIONARY fixtures).
• Full C++-to-Java audit: all 918 fixtures attempted; 846 passed, 72 failed (SPRINTZ fixtures).
• Failures reproduce across 32, 4,096 and 1,000,000 rows and all six compressors. No other failure categories were observed.
• Java/C++ Spotless checks, workflow YAML parsing, and git diff --check passed.

Independent fix validation

• [Java][CPP] Preserve SPRINTZ extreme transitions and block boundaries #1023 passes all 72 SPRINTZ combinations in each direction (four numeric types × six compressors × three scales), including 1,000,000 rows and deterministic 10% NULLs.
• With the regenerated SPRINTZ fixtures and both fixes applied, both full audits pass: Java-to-C++ 954/954 files, C++-to-Java 918/918 files. All original 108/72 failures are eliminated.
• The original audit results above describe the unfixed develop baseline.