fix(layout): don't panic collecting an empty stream

[miniex]

Summary

CollectStrategy collects its whole input into a single chunk for a child that requires exactly one chunk. When the input stream is empty, e.g. writing a zero-row table with a nullable struct column whose validity substream is empty, no chunk supplied a sequence id and it panicked with must have visited at least one chunk. CollectStrategy now yields nothing on empty input, and FlatLayoutStrategy returns an empty ChunkedLayout instead of requiring a single chunk, so no segment is written for an empty array.

The issue mentions the fuzzer's assume() guard could be dropped once this is fixed. I left it in place here: reading a nullable struct nested in a struct back is a separate bug (#8348), so the round-trip only works once both are fixed.

Closes: #8347

Testing

Added a regression test in vortex-file/tests/test_write_table.rs that writes a zero-row nullable struct column; it panicked before this change and now writes and reads back as zero rows. The issue's Python repro (vx.io.write of an empty struct column) no longer panics. cargo nextest run -p vortex-layout -p vortex-file passes (177 tests, including the segment-ordering tests). fmt --all + clippy --all-targets --all-features clean.

---

I'm Korean, so sorry if any wording reads a little awkward.

[codspeed-hq]

Merging this PR will not alter performance

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 3 improved benchmarks
❌ 7 regressed benchmarks
✅ 1571 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	decompress_rd[f64, (10000, 0.01)]	108.9 µs	139.3 µs	-21.82%
❌	Simulation	decompress_rd[f64, (10000, 0.1)]	109.2 µs	139.6 µs	-21.79%
❌	Simulation	decompress_rd[f64, (10000, 0.0)]	108.9 µs	139.2 µs	-21.77%
❌	Simulation	decompress_rd[f32, (100000, 0.0)]	496.1 µs	583.9 µs	-15.04%
❌	Simulation	decompress_rd[f32, (10000, 0.1)]	78.2 µs	91.4 µs	-14.4%
❌	Simulation	decompress_rd[f32, (10000, 0.01)]	78.2 µs	91.1 µs	-14.16%
❌	Simulation	decompress_rd[f32, (10000, 0.0)]	78.7 µs	91.3 µs	-13.88%
⚡	Simulation	bitwise_not_vortex_buffer_mut[128]	244.4 ns	215.3 ns	+13.55%
⚡	Simulation	bitwise_not_vortex_buffer_mut[1024]	304.7 ns	275.6 ns	+10.58%
⚡	Simulation	eq_i64_constant	317.9 µs	287.9 µs	+10.42%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing miniex:fix/empty-struct-write-panic (13b42fc) with develop (9814173)}

[onursatici]

Thanks for the contribution!
This seems correct, but I think we should fix this on a lower level.
We generally shouldn't use eof if we do not intend the chunk to go towards the end of file, so although it makes it possible to write an empty chunk, I think the right behaviour of collect strategy is to not yield anything on empty input.
If you just make that change the empty stream would propagate to the children and the flat layout would then fail. I think in the flat layout if there is no chunk we should return an empty ChunkedLayout which allows for no segments to exist at all, unlike flat layout which enforces one segment. If we do this we can save from writing a segment just to store an empty array

So the shape would be:

• CollectStrategy: empty input yields no chunks
• FlatLayoutStrategy: empty input returns a zero-row, zero-child ChunkedLayout
• non-empty flat input keeps the existing exactly-one-chunk behavior

[miniex]

Thanks for the steer - done.

CollectStrategy now yields nothing on empty input, and FlatLayoutStrategy returns an empty ChunkedLayout instead of requiring a single chunk, so no segment is written for an empty array. Dropped the eof minting as you suggested. Non-empty flat input keeps the exactly-one-chunk path.

Verified the empty file still opens and scans back to zero rows; full vortex-layout + vortex-file suites green.

[onursatici]

thanks, one final comment and then we can merge

[onursatici]

shall we invert this if and return early, something like:

  let Some(sequence_id) = latest_sequence_id else {
      return;
  };

  let collected = ChunkedArray::try_new(chunks, _dtype)?.into_array();
  yield (sequence_id, collected);

[miniex]

done, switched to the let-else guard.