fix(parquet): validate DataPageV2 decoded length

[fallintoplace]

What changed

• Make DataPageV2 V2 readers verify decoded byte counts instead of only checking buffer size.
• Update readV2Encrypted/readV2Unencrypted to return the actual number of bytes written/read and validate lengths.
• Enforce decoded-length validation in �6 uncompressed path and encrypted/unencrypted paths.
• Add regression tests for:
  • compressed DataPageV2 with mismatched codec decode length
  • encrypted uncompressed DataPageV2 with short decrypted output

Validation

• Added focused regression coverage in .
• No functional changes to non-V2 paths.

[zeroshade]

Thanks for adding the decoded-length validation — the goal is good, and the encrypted path plus the caller-level got != lenUncompressed check are correct. But readV2Unencrypted has an accumulation bug that regresses the happy path for compressed V2 pages (details inline). Requesting changes on that one check, plus a happy-path test to lock it in. Everything else looks fine.

[zeroshade]

This is the blocking issue. decodedLen += len(values) on the line above already makes decodedLen == levelsBytelen + len(values), so lenUncompressed - decodedLen here reduces to lenUncompressed - levelsBytelen - len(values) — which is 0 for any valid page. That turns this into len(values) != 0, i.e. it rejects every non-empty compressed DataPageV2. (The encrypted sibling gets it right: len(decoded) != lenUncompressed-levelsBytelen.)

Suggested fix:

decodedLen += len(values)
if decodedLen != lenUncompressed {
    return decodedLen, fmt.Errorf("parquet: metadata said %d bytes uncompressed data page, got %d bytes", lenUncompressed, decodedLen)
}

I verified on this branch with a valid compressed V2 page (Snappy, no levels): Next() returns false with the self-contradictory parquet: metadata said 128 bytes uncompressed data page, got 128 bytes. The new tests here don't catch it because they only exercise the mismatched and encrypted paths — please also add a happy-path test for a valid compressed DataPageV2, since the existing TestDataPageV2 is uncompressed and TestCompression is V1-only.

[zeroshade]

The compressed-V2 regression is fixed — readV2Unencrypted now checks decodedLen != lenUncompressed, and the new TestDataPageV2Compressed guards the happy path. Verified locally: that test passes (valid compressed V2 now reads, Data() round-trips) while TestDataPageV2CorruptDecodedLength and the encrypted-mismatch test still error correctly, and the full parquet/file package is green. Thanks for the quick turnaround — LGTM.