Point the differential gate at any released baseline

[derek73]

Summary

tools/differential/ answered one question — did the tree keep 1.4 parity — because it was built for the 2.0 migration. This generalizes it: --baseline VERSION compares the tree against any released version, so "what changes for a user already on the previous minor who upgrades and opts into nothing" is answerable at every release.

For 2.1 that answer is 89 diffs, every one on a name bearing a CJK character, zero on Latin-only input.

--baseline 1.4.0 → 751 names; intentional diffs: 107; unexplained: 0; 12 of 107 Latin-only
--baseline 2.0.0 → 751 names; intentional diffs:  89; unexplained: 0;  0 of  89 Latin-only

Changes

• --baseline VERSION (default 2.0.0). The version selects the ledger (expected_since_<VERSION>.toml, a hard error if absent — not an empty rule set, which would classify nothing and read as a catastrophic regression) and the compared surfaces: facade alone below 2.0, facade + v2 API + AmbiguityKinds from 2.0 on.
• The worker is generated per run with its PEP 723 pin substituted, into a temp dir outside the worktree. That placement is the safety mechanism, not plumbing: uv reads genuine inline metadata, and sys.path[0] becomes a directory holding no nameparser, so the checkout cannot shadow the pinned wheel. worker_v1.py is deleted.
• The worker emits a version tell, verified before any comparison. The README argued for this at length; it had never been implemented, leaving pin verification a manual step nobody performed. Both halves are checked and neither implies the other — an editable install reports the tree's version, so version agreement proves nothing when tree and baseline share a number.
• Ledger fields name roles the way Role does. The facade's first/last canonicalize on the way in, and the UNEXPLAINED block prints canonical names so what a reader sees is what they write in a rule. AGENTS.md already makes Role canonical "defined once and derived everywhere"; the facade's vocabulary expires at 3.0.
• expected_changes.toml → expected_since_1.4.0.toml (rename kept 100% similarity, so git log --follow still reaches the file's origin), plus the new expected_since_2.0.0.toml.
• New tests/v2/test_differential.py — 38 tests over the logic whose silent misbehavior produces false confidence.
• AGENTS.md drops the corpus count and file roster rather than correcting them. The line was stale at "~650 names" and "two files" despite already carrying a hedge explaining why it gave no exact count — and the costly half was never the count but "two files", which hid corpus_cjk.jsonl, the corpus covering the whole 2.1 CJK surface.

Known ceiling, documented in the README

The corpora run under the default policy, so policy-gated behavior is invisible here. #329 is the worked example: under Policy(maiden_delimiters={("(", ")")}) the tree gives maiden Jones where 2.0.0 gives née Jones, and this gate cannot see it. Default vocabulary is fully exercised — Lexicon.maiden_markers ships 17 entries and the bare née Jones form needs no configuration. What is opt-in is the delimited path, not the marker words. tests/v2/cases.py covers opt-in behavior via its per-row policy=.

Review found one hole worth naming

The spaced-honorific rule matched a leading 양/군 — ordinary surnames in that position (Yang, Gun). Nothing was misclassified today, but a future change routing the leading 양 of 양 미선 to suffix would have fallen through to that rule, whose fields include suffix, and been reported as an intended honorific. A surname silently reclassified as a suffix is precisely the regression class this gate exists to catch. Head match dropped; all 89 still classify.

Verification

• uv run pytest -q → 3013 passed, 20 skipped, 11 xfailed
• uv run mypy → clean, 97 source files
• Both baselines exit 0, both tells confirmed loading from the uv cache rather than the checkout
• The UNEXPLAINED path exercised against a stub ledger, confirming it prints canonical field names

Follow-up left open

tests/v2/test_regex_sync.py pins the 1.4 ledger's hand-copied character classes against their sources by hardcoded filename, so the 2.0 ledger's copies are unpinned. Both were verified equal to their sources today, and the gap is recorded in the ledger's own comments. Parametrizing those pins over every expected_since_*.toml needs the #271/#272 selector reworked first, since it asserts uniqueness and the new file has two such rules.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.48%. Comparing base (88ca0b3) to head (b30f174).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #332   +/-   ##
=======================================
  Coverage   98.48%   98.48%           
=======================================
  Files          41       41           
  Lines        2832     2832           
=======================================
  Hits         2789     2789           
  Misses         43       43           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.