chore(deps): bump github.com/lanrat/extsort from 1.4.2 to 1.5.0

[dependabot]

Bumps github.com/lanrat/extsort from 1.4.2 to 1.5.0.

Release notes

Sourced from github.com/lanrat/extsort's releases.

v1.5.0

Highlights

This release fixes several serious bugs in the sort pipeline: a hang, silent data loss, and crashes caused by user callbacks. It also makes sorting about 2× faster. Upgrading is recommended for everyone.

⚠️ Breaking changes

• Go 1.25 or newer is required (previously 1.23), because of the golang.org/x/sync v0.22.0 upgrade.
• diff.NEW and diff.OLD are now typed diff.Delta constants. Code that uses them as a plain int, such as var d int = diff.NEW, no longer compiles. Use diff.Delta, or convert with int(diff.NEW).

Behavior changes

• Errors that were lost are now reported. Read errors during the merge, panics in a legacy FromBytes, and temp-file Close errors now arrive on the error channel. If you were getting incomplete output with a nil error, you'll now get an error instead.
• Callback panics no longer crash the process. A panic in your compare, fromBytes or toBytes function during the save or merge stage is returned as a ComparisonError, DeserializationError or SerializationError.
• An unusable explicit TempFilesDir is an error again. A path that is a regular file, or that has no write permission, now fails instead of silently falling back to /var/tmp. This restores the behavior from before v1.4.1. A directory that doesn't exist yet is still created.
• The temp file is created only when needed, once the input needs more than one chunk. Temp-file errors arrive on the error channel, and the constructors never return a nil sorter.
• ChanBuffSize now defaults to 1 (previously 16), matching the docs and using less memory. A zero value still means unbuffered, and the docs now say so.
• Concurrency is documented. Your compare, fromBytes and toBytes functions may be called from several goroutines at once. The legacy New() / NewMock() API now serializes calls to your FromBytes, restoring the thread safety it had before v1.1.0.
• Default temp-dir discovery creates and removes one probe file per candidate directory, once per process.
• diff with an already-cancelled context now returns immediately.

Bug fixes

• Sort() no longer hangs forever when saving a chunk fails, for example on a full disk.
• A read error on a chunk's first record no longer silently drops the chunk.
• Closing the temp file can no longer panic with "send on closed channel".
• Small, empty and failed sorts no longer leak temp files and file descriptors. On Windows, temp files are now removed.
• Default temp-dir selection falls back correctly when /var/tmp is missing or read-only, for example under Kubernetes readOnlyRootFilesystem. This was a regression in v1.4.1.
• Sorting no longer spins for up to 2×ChunkSize iterations after the input closes. A 10-record sort went from about 50 ms to microseconds.
• Several sorters can now share one *Config without a data race. The caller's struct is no longer modified.
• diff no longer hangs when an error channel is never closed, and it respects context cancellation.
• UniqStringChan now accepts the <-chan string returned by Strings().

New

• diff.StringResultChanContext(ctx) works like StringResultChan but stops when ctx is cancelled.

Performance

• Records move between merge workers in batches instead of one at a time.
• Ordered uses a binary codec instead of creating a gob encoder for every record.
• Strings and Ordered reuse their encode and decode buffers.
• Merge read buffers are allocated lazily and sized to each chunk, so memory no longer grows with the chunk count (2,000 chunks: 137 MB → 7 MB).
• Chunks grow as needed instead of allocating ChunkSize up front.
• The unneeded fsync on the temp file is removed.

Benchmarks, Apple M5 Max:

Benchmark	Before	After

... (truncated)

Commits

• a49dcec test: detect temp files in the single-chunk optimization tests
• cffc0e8 release: refuse to tag a dirty, stale or non-main tree
• 0ff3472 Merge pull request #15 from lanrat/claude/extsort-perf-batching
• dab2261 Reuse the encode and decode buffers of the built-in codecs
• 9ce531a Grow chunks instead of allocating ChunkSize up front
• 925c333 Receive input without locking ctx's channel for every record
• 07d4eec Batch records between the parallel merge workers and the final merge
• 60595a6 Remove the unused byteSlicePool
• 584a942 diff: receive without locking ctx's channel for every value
• 2129eb8 queue: keep elements in a heap of values
• Additional commits viewable in compare view