fix(foundation): extracting from a pipe or non-seekable stream silently stops early (870 of 1,117 frames) #1586

Reading a capture from a pipe or other non-seekable stream drops frames without a warning. The default engine stops partway through.

Repro on main (e9f3340f7), using examples/captures/http.pcap (1,117 frames):

• extract(fin='…/http.pcap') reads 1,117 frames.
• cat http.pcap | python -c "…extract(fin=sys.stdin.buffer, nofile=True)…" reads 870, both with and without no_eof=True.
• An os.pipe() written in 4,096-octet chunks also reads 870.
• A non-seekable io.RawIOBase that returns short reads also reads 870, both raw and wrapped in BufferedReader.

This is the input pcapkit - (stdin) relies on: __main__ sets no_eof exactly for that case.

Likely cause, not yet traced: a short read() from a pipe returns fewer octets than requested, and something treats that as end of file. A pipe may return less than requested at any point. The #1569 review saw the default engine read only 56 of 1,117 frames with a different chunking, so the cut-off depends on chunk boundaries.

Expected: every frame is read from a non-seekable stream, the same as from a path. A short read must not end the capture.