feat: C-library interop testbench + 6 write-compatibility fixes

[rhaist]

Summary

Adds an interop test suite that validates Go-written files against the official HDF5 C tools (h5dump/h5diff/h5repack), plus a c-interop CI job (Ubuntu, hdf5-tools). Tests self-skip when the tools are absent, so the existing jobs are unaffected.

• TestCInterop_WriteMatrix — 15 cases across the write feature matrix (all numeric types, multidim, chunked+gzip, shuffle+fletcher32, fixed/vlen strings, ragged vlen, groups+attributes, hard/soft links, enum, array types, superblock v0), each verified through real h5dump -p output including data values.
• TestCInterop_RepackRoundTrip — Go write → h5repack (C rewrites) → h5diff (C compares) → Go reads both files back and verifies every value.

Bugs the suite caught — all fixed

1. Deflate filter wrote gzip (RFC 1952) instead of zlib (RFC 1950) — every compressed dataset was unreadable by the C library, h5py, and our own Open() path. The reader keeps a gzip-magic sniff fallback so files written by earlier releases stay readable.
2. Filter pipeline message: version 2 stamp with a version 1 body — h5dump parsed filter ID 0 ("unknown filter"). Now emits spec-correct v1 (padded null-terminated names, 4 pad bytes after odd client-data counts).
3. Enum datatype message interleaved name/value pairs — the spec stores all names, then all values (0 length enum name in h5dump). Version stamp corrected v3 → v1 to match the padded layout.
4. Hard links: RefCount message missing its version byte + EOA left 8 bytes short — fixed the message layout (reader accepts the old bare-uint32 layout for back-compat) and routed header rewrites through the bounds-checked writer; delete_write.go's refcount path got the same fix.
5. Soft links written as standalone object headers holding a Link message — the C library cannot classify such an object. Now written per spec: cache-type-2 symbol table entry with the target path in the parent's local heap. SNOD entries now also serialize scratch-pad data they previously zeroed (cached STAB addresses were silently dropped on rewrite).
6. In-place header rewrites could overflow their allocation — a dataset header growing past its 256-byte allocation overlapped the following GCOL by 2 bytes, corrupting the header checksum. Dataset header allocations are now tracked (headerAllocs) so the existing OCHK continuation path actually fires.

Also: the Windows-hardcoded TestVLenStringAttribute_H5dump now uses shared helpers and runs everywhere (renamed TestCInterop_VLenStringAttribute); stale "superblock v3 not implemented for writing" doc comments corrected (v3 writing works and is covered by the suite).

Test plan

• go test -run TestCInterop -v . with HDF5 2.1.1 tools: 16/16 pass, no skips
• Full suite: 3712 pass across 18 packages
• New CI job runs the interop tests on every push/PR

Follow-ups (not in this PR)

• HDF5 2.x "latest"-format read gaps: layout-v4 chunk indexes, complex datatypes, float16
• FP8/bfloat16 helpers exist but are not wired into the read path
• README compatibility claims need tempering until the above land

[codecov]

Codecov Report

❌ Patch coverage is 91.66667% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/writer/filter_gzip.go	57.14%	3 Missing ⚠️
group_write.go	83.33%	1 Missing and 1 partial ⚠️
link_write.go	71.42%	0 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!

[kolkov]

Enterprise-level review complete. All 6 bug fixes verified against HDF5 C reference source (H5Zdeflate.c, H5Opline.c, H5Odtype.c, H5Orefcount.c, H5Glink.c/H5Gent.c, H5Oalloc.c/H5Ocache.c).

Fix 1 (Deflate): Correct — C uses compress2() = zlib RFC 1950, not gzip. Gzip sniff fallback mathematically safe.
Fix 2 (Filter pipeline): Correct — v2 stamp + v1 body was causing 'unknown filter' in h5dump. Now proper v1.
Fix 3 (Enum): Correct — interleaved layout caused '0 length enum name'. Now all-names-then-all-values per H5Odtype.c:1329-1349.
Fix 4 (RefCount): Correct — missing version byte (H5Orefcount.c requires 5-byte layout). EOA fix via writeOHDRWithBoundsCheck.
Fix 5 (Soft links): Correct — standalone OHDR was fundamentally wrong for symbol table format. Cache-type-2 per H5Glink.c:374-383. Bonus: SNOD scratch-pad fix prevents data loss on rewrite.
Fix 6 (Header overflow): Correct — headerAllocs + bounds check + OCHK continuation matches C library pattern (H5Oalloc.c).

Backward compatibility verified for all fixes with tested fallback paths. No regressions. Excellent work.