[ShiTan-677]

When a name matches two to five contacts, read_chat() reads every matched conversation and the export CLI silently selects the first. Only more than five matches were rejected, despite the existing Skill/MCP guidance to ask for a selection.

Reject every non-unique match before fetching message contents or opening an output file. Return/display candidate wxids so callers can retry with an exact identifier. The no-match and single-match behavior remains supported.

Validation on macOS / Python 3.12:

• Before the fix, the read/export regressions failed for two and five matches.
• python -m pytest -q: 91 passed, 1 skipped.
• Synthetic tests cover two, five and six candidates, exact-identifier selection, no match, and preservation of an existing export file when selection is ambiguous.
• Compilation, macOS shell syntax and six-tool MCP loading checks passed.

Export-selection fixtures are uncompressed, so they stub only the optional zstd availability gate. Modern timestamps and explicit UTF-8 keep the synthetic exports portable across macOS and Windows.

This PR is independent of the sender-direction fix. It contains only synthetic contact fixtures, without personal chat data or keys. The final fork CI matrix passed on macOS and Windows with Python 3.10 and 3.13: run 38039930827, at commit 80942a0. The upstream fork-PR workflow still requires maintainer approval.

[freanocank]

Thanks for the fix and regression tests. Adopted through #15 with your original commits preserved. Ambiguous names now require an exact contact selection before message reads or export writes.

The combined suite passes all four macOS/Windows × Python 3.10/3.13 CI jobs; Windows reports 209 passed, and macOS reports 208 passed with the Windows-only test skipped.