Parquet corpus from third-party writers #35

Part of #22. Independent of everything else.

Add a small corpus of parquet files written by tools other than the Arrow Rust writer, and run the existing local tests plus the fuzz harness against them.

Why a corpus rather than a generator

#10 was a file-shape bug: the plan-time opener used PageIndexPolicy::Required, so a scan failed outright on valid parquet where a column chunk advertises a page index but carries no offset index. The Arrow Rust writer does not readily produce that shape, so a generator built on it is unlikely to find the class — but files in the wild have it, which is how the bug was found.

This is the honest answer to the input-diversity dimension, and it is cheaper than teaching a generator to emit deliberately unusual physical layouts.

dev/test_parquet/openobserve.parquet is already in the repo and is exactly this kind of artifact — a file from a third-party writer that exercises paths the generated files do not. It is used by two tests. The gap is that there is one of it.

Scope

• Files from pyarrow, duckdb, and spark, plus any real-world file that has previously broken a scan
• Varied physical shapes: missing offset indexes, unusual page layouts, nested schemas, legacy writer versions, mixed encodings
• Keep them small — this is about shape, not volume — and document each file's provenance and what makes it interesting
• Run them through the fuzz harness's differential oracle, not only through fixed tests

Acceptance

• A regression file for fix(reader): treat the parquet page index as optional #10's shape specifically, so that class cannot silently return.
• Each corpus file has a one-line note explaining why it is in the corpus. A corpus of unexplained binaries rots.