fix: module compression incorrectly reused first file's root pointer

[kim-em]

Summary

Fixes a bug in the module compression feature where .olean.server and .olean.private files were being truncated to 96 bytes (just the header) instead of being fully compressed.

Note: This diagnosis and fix were performed by Claude Code (@anthropics). While integration tests pass, this needs careful review by @digama0 to ensure correctness, as I (@kim-em) cannot vouch for the implementation details.

The Bug

Two related issues in src/lgz.rs compress() function:

1. Line 261: Used root instead of header2.root when storing each file's metadata

   • Caused all Data entries to reference the same root object from the first file

2. Line 279: Used root instead of d.root during reference counting

   • Only marked the first file's root as referenced

Symptom

When compressing three module files (.olean, .olean.server, .olean.private):

• ✅ First file (.olean): Compressed correctly → ~48KB
• ❌ Second file (.server): Decompressed to 96 bytes instead of 2.3KB
• ❌ Third file (.private): Decompressed to 96 bytes instead of 46KB

The second and third files were truncated because the compressor's backref cache thought it had already written their root objects (since all three files pointed to the first file's root).

The Fix

Two one-line changes in src/lgz.rs:

1. Line 261: root → header2.root
2. Line 279: root → d.root

Each file now correctly stores and compresses its own root pointer, producing three complete, distinct object graphs.

Testing

#4 added integration tests in tests/module_compression.rs with test data from Plausible.Attr module.

Before fix:

✓ test_single_olean_roundtrip: PASSED
✗ test_module_roundtrip: FAILED
  - Attr.olean.server: 2344 bytes → 96 bytes (truncated!)
  - Attr.olean.private: 47320 bytes → 96 bytes (truncated!)

After fix:

✓ test_single_olean_roundtrip: PASSED  
✓ test_module_roundtrip: PASSED
  - Attr.olean:         48976 bytes ✓
  - Attr.olean.server:  2344 bytes ✓
  - Attr.olean.private: 47320 bytes ✓

cc @digama0 for review

[kim-em]

(If anyone is interested in what the Claude conversation looked like, it's here.)