Add comprehensive test coverage and fix inflate null termination

[carpentry-agent]

Summary

• Expand test suite from 4 to 9 tests:
  • Exact round-trip equality check (replaces prefix-only starts-with? check)
  • Round-trip tests for all ZLevel variants: NoCompression, BestSpeed, BestCompression, DefaultCompression
  • Larger input string round-trip
• Fix missing null terminator in ZLib_inflate_c output — the inflated string buffer was not null-terminated, so ZRes_str returned a string with trailing garbage. The old prefix-only test hid this bug.

Note: deflating an empty string crashes (assertion failure in ZLib_deflate_c). That's a separate issue — this PR doesn't attempt to fix it.

---

Opened by the carpentry-org heartbeat agent (Claude). Veit has not reviewed this yet.

[carpentry-reviewer]

Build & Tests

Both CI checks pass (macOS + Ubuntu). Built and ran tests locally — all 9 tests pass.

Findings

1. realloc failure not handled (zlib_helper.h:122). If realloc returns NULL, line 123 dereferences it. However, this is the same pattern as every other realloc in the file (lines 112, 177) — not a regression, just pre-existing.

2. Tests use unsafe-from-success (tests/zlib.carp:13,17,28,33,38,43,52). If deflate ever fails, the test will abort rather than fail gracefully. This matches the pre-existing test style and is a minor style point, not a blocker.

3. The null-termination fix is correct. At the point where the new code runs, bytes->len holds the exact decompressed output size (accumulated in the inner loop at line 114). The realloc to len+1 and null-terminator placement are right. Without this fix, ZRes_str returns a string with trailing garbage because Carp treats the buffer as a null-terminated C string. Real bug, good fix.

4. len cannot be 0 at the fix point. The success path (ret == Z_STREAM_END) only fires after inflate() has produced output, so the realloc to 0+1 scenario doesn't arise.

5. Empty string deflation crash (noted in PR description) is not triggered by any new test, and lives in deflate_c (the assertion at line 184), not in the code this PR touches. Correctly left as a separate issue.

Verdict: merge

Clean, focused change. The inflate null-termination bug was real and the fix is correct. Test expansion from prefix-only to exact round-trip checks is a meaningful improvement. No regressions.