Fix joiner-aware text normalization and version MNG canonical output

[satsragclaw]

Summary

• use one Mongolian word-run classifier for strict validation and normalize_text(), preserving Nirugu/ZWJ joining context
• freeze the current MNG canonical selection policy as mng-canonical/1, expose it through MongolianShaper.canonical_version, embed it in the normalize table, and reject mismatched table versions
• add reproducible compatibility fixtures:
  • 15 ordered phase-trace vectors covering every MNG shaping rule transition
  • 1,993 byte-for-byte canonical vectors covering every current corpus shape group
• add scripts/gen_compat_goldens.py with write and --check modes; serialization is key-sorted for Python 3.6 portability
• fail on unknown corpus aliases, freeze the 1,993-vector cardinality, and require every phase vector to trigger its declared witness rule
• include test data, goldens, and generator scripts in source distributions

Why

normalize_text() previously split runs at Nirugu and ZWJ even though normalize() accepted both as word characters. Splitting at a joiner changes neighboring positions and could produce a different canonical form.

The goldens lock the current shaping transitions and canonical Unicode outputs so later refactors cannot change them silently. They were replayed against unmodified upstream/main@8d43521962ef489443e18f900491ba9f5ea72e05 with zero differences.

Verification

• python3 -m compileall -q mongol_norm scripts tests
• python3 scripts/gen_compat_goldens.py --check
• python3 -m unittest discover -s tests -p 'test_*.py'
  • 157 tests passed
  • 177/177 core-hud direct cases
  • 3507/3507 eac-hud direct cases, 5 documented UTN xfails skipped
  • 225/225 CORE-HUD round trips
  • 3513/3513 EAC-HUD round trips
  • 2237/2237 prefix-stability pairs
  • 1993/1993 shape groups converge
• normalize-table regeneration SHA-256 unchanged across repeated generation
• uv build succeeded for wheel and sdist
• full 157-test suite passed from the unpacked sdist
• installed-wheel smoke test passed on Python 3.14
• unmodified-base replay: 1993 canonical vectors, 0 differences; 15 phase vectors, 0 differences

Closes #10

[satsragclaw]

Superseded by #13 and #14 so the joiner bug fix and the canonical-version/compatibility work can be reviewed and merged independently. The final content is preserved across those two PRs.