comparator: pair columns by agreement, an unordered door, the outputs claim (ACE-131)

[sandeep-agami]

Spec: ACE-131 (agami-sdlc projects/agami-core-extension/specs/reconcile-evidence/ACE-131-comparator-agreement.md)

Stacked on #315 (ACE-130). Round 9 of the reconcile stack, from the 51-row run.

The findings

1. Order sensitivity was wrong for reconcile. 2e handed the person's statement as --golden-sql-file, has_top_level_order_by set ordered=True, one reordered row unpaired every column, the early return fired before the values level's forgiveness, column_pairs came back empty and every safety net in reconcile.py is keyed on pairs. Two rows were false mismatches on sort order alone.
2. All-or-nothing pairing over-reported. One differing row unpaired the column and the page printed "no generated column carries the values of: …" for a column that was there; four rows had real differences the report buried.
3. "Can the reconcile skill determine if the query is equivalent based on the join and filters?" Partly: the seven claims never covered the projection.

The change

comparator.py

• pair_columns: stage one is today's exact pairing (names and positions never consulted). Stage two, over what is left: a same-named column (lowercase, qualifier dropped) pairs at any agreement; otherwise the candidate with the highest share of agreeing rows pairs when the share is strictly above one half. Agreement is positional when ordered, a multiset when not, over canonical cells with the level's quantize. match_columns stays a two-tuple wrapper, so every caller keeps its shape.
• ItemScore gains column_agreement (pair for pair) and paired_row_share (the share of rows every paired column agrees on, computed even when a golden column has no partner). Additive, defaults.
• compare_result_sets(..., ordered=None): when given it decides instead of the statement; False adds the note "row order was not compared". The golden run never passes it.
• The pass mark does not move: an item passes at exactly 1.0; a golden column with no partner still scores 0 with its name. What changes is the failing item's number and reason ("9 of the answer key's 10 rows matched").

golden_claims.py: the eighth claim outputs, after tables: one key per output expression through _expression_key, alias peeled; SELECT * and t.* read *; DISTINCT is not read. Reports, never gates. Docstrings, sm claims help and the plugin prose that counted seven count eight.

cli.py: sm compare-results --unordered.

Pins moved on purpose

• test_column_values_in_a_different_row_order_do_not_match_when_ordered → the column pairs by name and compare_rows finds 0 of 2 rows in order; the score reads the rows reason, not the missing-column one.
• test_a_bool_column_does_not_match_an_int_column and the end-to-end boolean test → paired by name at agreement 0.0, accuracy still 0.0.
• test_a_null_does_not_match_the_empty_string_end_to_end → paired at agreement 0.0, accuracy still 0.0.
• test_each_level_over_a_passing_and_a_failing_pair[exact …] → 0.5 (one of two rows agrees) instead of 0.0.
• test_matching_forwards_quantize → the name-free case keeps the old assertion; the same-name case checks agreement 0.0 against 1.0.
• test_the_claim_set_is_exactly_seven_claims → eight, with outputs; test_a_scored_item_carries_the_seven_claims likewise; two * 7 status lists → * 8.

Tests

New: a column agreeing on 9 of 10 rows pairs and scores 0.9; a differently named column agreeing on 3 of 10 stays unmatched with paired_row_share 1.0 over the paired one; a same-named column pairs at any agreement; ordered=False against an ORDER BY statement scores 1.0 with the note, and the statement still decides when the caller says nothing; differing row counts leave pairs, agreement and share empty; --unordered at the CLI; the outputs claim (aliases peeled, a different aggregate differs, a bare column in a single-table statement is its table's, SELECT * differs, never gates). 639 passed across the comparator, claims, CLI, golden and reconcile suites (the golden_run client test is the known local failure).

🤖 Generated with Claude Code