[smounir]

Fix OOM in extract when the source is a local file

Problem

FileBucket.NewRangeReaderEtag allocated a []byte of the full requested range length and filled it with ReadAt. This makes extract allocate the entire merged range in memory.

With --overfetch=0 on a contiguous zoom level, mergeRanges coalesces adjacent tiles into very large ranges. On a 2.9 TB source (--minzoom=18 --maxzoom=18), one of the 254 requests was ~218 GB, so makeslice failed:

fatal error: runtime: out of memory
...
pmtiles.FileBucket.NewRangeReaderEtag  bucket.go:136
pmtiles.Extract.func2                  extract.go:538

GOMEMLIMIT doesn't help, because this is a single allocation. The HTTP and cloud bucket implementations already stream the response body, so only FileBucket was affected.

Fix

FileBucket.NewRangeReaderEtag now returns an io.SectionReader over the open file, wrapped in a ReadCloser whose Close() closes the file. Memory use no longer depends on the range size.

• I also removed the defer file.Close(). It was registered before the error check and would have closed the file before any streaming read.
• Behavior is otherwise unchanged:
  • 404 if the file is missing.
  • 412 if the etag doesn't match.
  • 206 otherwise.
  • A range running past EOF is truncated, as before.
• The old 416 branch ("Expected to read N bytes...") was unreachable, since ReadAt always returns an error when it reads fewer bytes than requested. It's removed.

Tests

Added TestFileBucketRangeReader, which covers:

• a range in the middle of the file
• a range truncated at EOF
• an offset past EOF
• a missing key
• a stale etag (412)