Report the evaluated value when a predicate application violates a prohibition or requirement

[rmichaelthomas]

Summary

• _condition_actual_values (interpreter.py:1915) handled ConditionNode/CompoundConditionNode but fell through to "" for PredicateApplicationNode. _exec_forbid/_exec_require only append the result when non-empty, so forbid <field> is <predicate> named the field and the predicate but never the value that tripped it — e.g. a Section 42 income-limit violation said "household-income-at-move-in is initial-income-over-limit." with no mention of the $65,000 that actually triggered it. Equality-style conditions already reported their value; this was a gap, not a design choice.
• Adds a PredicateApplicationNode branch: reports the subject's value plus the predicate's own body — household-income-at-move-in is 65000. initial-income-over-limit: is above 58320.
• Judgment call (flagging per the build prompt): a predicate body's field is an implicit EachPronoun, so rendering it directly produces each is above 58320 — "each" reads as noise in a violation message, so I strip a leading each before appending. This is a formatting choice, not a correctness one; happy to go the other way if you'd rather see it verbatim.
• Scoped to exactly one function. _exec_forbid/_exec_require are unchanged (they already append actual when non-empty). render()'s output for every existing node type is unchanged — confirmed by the full existing suite passing, including round-trip tests.
• Bumped pyproject.toml to 0.18.1 (bugfix in message content, not a language or API change).

Pre-branch confirmation (B.0)

Confirmed HEAD (de96e45) is exactly what PyPI serves as 0.18.0 before touching anything: downloaded the real liminate-0.18.0.tar.gz sdist from PyPI and diffed its src/liminate against this repo at de96e45 — byte-identical except __pycache__. Also confirmed 5db033d..de96e45 shows the TI-Q15 commit (#69) itself bumped pyproject.toml to 0.18.0, so there's no unexplained gap. No divergence to report.

Tests

Added 8 tests to tests/test_define.py, all smoke-tested against the running fix before being written down:

1. forbid <field> is <predicate> violating → message contains the value and the predicate body
2. require <field> is <predicate> not met → same
3. Predicate applied to an unset subject → returns "", nothing raises (direct call, mirroring the pre-existing ConditionNode test shape)
4. is not <predicate> (negated) violating → value still reported
   5/6. Compound condition (A and B) → reports whichever branch actually failed, whether it's the ConditionNode or the PredicateApplicationNode side (two tests, one per direction)
   7/8. Regression, blocking: exact byte-for-byte pre-existing message text for plain ConditionNode forbid/require violations, unchanged

Verification

• pytest tests/ — full suite: 1859 passed, 0 failed

• pytest tests/test_define.py — 55 passed (44 pre-existing + 11 new, including 3 parse-only predicate tests untouched by this change)

• Local end-to-end check against the real Section 42 evidence-noncompliant.json (household-income-at-move-in: 65000) through the fixed interpreter and the boundary-corrected agreement (liminate-dev PR #100) produces exactly:

  Prohibition violated: household-income-at-move-in is initial-income-over-limit. household-income-at-move-in is 65000. initial-income-over-limit: is above 58320.

  This is a local sanity check, not the live re-run — that happens after publish + the liminate-dev pin bump (B.6), against the real Agreements API.

Test plan

• pytest tests/ — 1859 passed
• Local end-to-end check against real Section 42 evidence — exact expected message
• Rob: merge, then publish 0.18.1 to PyPI
• Rob: bump the liminate pin to >=0.18.1 in liminate-dev, deploy, confirm live
• Then: re-run the §42 noncompliant fact pattern live, save receipt-noncompliant-v2.json, confirm the income-limit rule's message contains 65000

🤖 Generated with Claude Code