…undtrip

Adds tests/test_export_import_roundtrip.py with four hermetic ONNX-backed
baselines for cross-instance export/import behaviour. The tests do not try
to assert a future desired state — they pin down what the current impl
actually does today, so Phase 2 has measurable targets.

Why:
  The mm export/import plan wants cross-PC roundtrip fidelity plus additive
  ingestion. Before designing the fix, we needed concrete numbers for what
  breaks today. This file is that measurement.

What the baselines show:
  * Single cross-PC roundtrip: content/metadata/top-k search are fully
    preserved; only chunk UUIDs are reassigned on import.
  * Re-importing the same bundle creates exact 2x row duplication
    (4 contents -> 8 rows, max 2 rows per content_hash) because
    upsert_chunks dedupes by UUID and import assigns a fresh UUID.
  * Disjoint merge works (A+B rows, no hash collisions, B's native
    content stays searchable post-import).
  * Content-collision merge also duplicates: identical content on both
    sides produces two rows with the same content_hash but different UUIDs.

Two of the four tests hard-assert the current buggy behaviour (2x dup on
re-import, duplicate row per collision). They are intentional regression
markers: when Phase 2 introduces content_hash-based dedup, these must fail
and be flipped to assert the corrected counts.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

The Phase 1 baseline previously printed metadata and top-k overlap as
diagnostic output without enforcing them in hard asserts. The PR body
claimed "content / metadata / top-k search fully preserved" but only
the content-hash equality was automatically guarded, so a future
regression in metadata wiring or embedder drift would silently slip
through with just a print delta.

Promote both to hard asserts in test_phase1_baseline:
  * metadata mismatches across tags, namespace, heading_hierarchy, and
    source_file must be empty (pair by content_hash).
  * top-3 result sets for all four probe queries must match exactly
    between the source and the re-imported instance (set equality,
    order-insensitive to tolerate any tie-breaking noise).

Verified locally: overlap=3/3 for all queries, metadata mismatches=0.
Claim in the PR body now matches what the test actually enforces.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>