test(semantic): raise experts.py mutation kill rate 44.4% -> 63.2%

[kevincostner17]

What

semantic/experts.py is the module where semantic repairs are proposed, and it killed only 111 of 250 mutants. This adds 60 tests across four themes.

Module	Before	After	Survivors after
semantic/experts.py	44.4% (111/250)	63.2% (158/250)	92 (untargeted)
semantic/scoring.py	96.6% (56/58)	98.3% (57/58)	1, proven equivalent

Of the 37 targeted mutants: 32 killed, 5 proven equivalent, 0 unexplained.

All runs cover every mutation site with not applied: 0.

Why the survivors mattered

Expert applicability guards. if info.free_text or info.identifier_like or info.boolean_like: could become and, and info.numeric_like and not info.free_text could become or, with nothing failing. These guards are what keep an expert off identifier and free-text columns — the brief's "ID-protection removed" class. An expert that runs on an identifier column rewrites "007" to 7, and the damage is unrecoverable from the output alone.

Date day/month disambiguation. The a > 12 and b <= 12 family decides whether 05/12 is May 12th or 5th December. Every boundary at 12 was movable — a silent wrong date written into user data.

Currency and number parsing, including the ambiguity flag for "1,000" with no currency code — the corpus trap this parser exists to handle — and the decimal-separator selection.

Allowed values and the category tie-break, plus the frozen-ness of _DateResolution and dropna in the value counter.

The five equivalent mutants are proven, not asserted

• cmp#7 — reachable only inside if dots and commas:, so both rfind calls return ≥ 0 and, being different characters, can never be equal; > and >= agree on every reachable input. Confirmed over 727,831 parser calls.
• const_bool#5 — _valid_grouping line 134 is dead: an early return guarantees rest is non-empty. Confirmed by sys.settrace line coverage over 610,436 calls: zero executions.
• cmp#20, cmp#21, bool#16 — the three branch-3 date mutants differ from the original only in states already claimed by branches 1 and 2, which return first. Confirmed by differential over the complete input domain (a, b ∈ 0..99, both separators, 3 years, 3 dayfirst values) = 180,000 inputs per mutant, zero differences.

This also corrects PR #491

#491 reported 98.3% for scoring.py. That figure was wrong. Its verdicts came from a harness run affected by a stale-bytecode defect, and the true value at that commit was 96.6%.

The masked survivor was json.dumps(features, sort_keys=True) → False in features_hash — the property that makes the digest a function of feature content rather than of the order the dict literal happens to be written in. Every existing test compared hashes built in a single order, so none of them noticed. A test for it is included here, and scoring.py now genuinely measures 98.3%.

Scope

Tests only — no src/ change. No behaviour changes, so no changelog entry and no compatibility impact.

Verification

• The four new files + test_scoring_boundaries.py — all pass
• Full suite, py3.12 — 7309 passed, 22 skipped, 0 failed, coverage 95.01%
• ruff check + ruff format --check clean
• Mutation re-runs as tabled above

Two defects found in the library (reported, not fixed here)

Neither is fixed in this PR, because both touch src/ and warrant their own change with its own regression test.

1. The currency parser's ambiguous flag is dead on every path. experts.py:110-114 documents that "a currency absent from it and not settled by structure is reported ambiguous rather than guessed". It cannot happen: _split_amount's only caller always resolves a currency code first, so the ambiguous branch is unreachable (verified: 0 of 98 public inputs), and parse_currency — the entry point all six consumers use — discards the flag anyway. No recognised currency produces a wrong value today, since all eight codes outside {EUR, CHF} are dot-decimal locales. The hazard is latent: adding SEK/BRL/DKK to _CURRENCY_CODES without also adding it to _COMMA_DECIMAL_CURRENCIES would read SEK 1.200 as 1.2 instead of 1200, and the documented safety net would not catch it.

2. The category tie-break is decided by row order — a metamorphic violation. On an exact tie, count > best[1] keeps whichever spelling value_counts saw first, which is a property of row ordering, not of the data:

input order	proposal at defaults	data at semantic_auto_threshold=0.90
["USA"]*3 + ["usa"]*3	'usa' → 'USA'	all six become USA
["usa"]*3 + ["USA"]*3	'USA' → 'usa'	all six become usa

At defaults the data is unchanged (confidence 0.93 < 0.95) but the proposal direction still flips, so the report is order-dependent even at defaults. semantic_auto_threshold is a documented public kwarg, so the data flip needs no private API. An exact tie means there is no column-local evidence for either spelling, so the contract says route to review rather than pick one. The tie-break tests in this PR are characterization tests whose docstrings say plainly that they pin a suspected defect, so a deliberate fix has to come past them.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 516b00c8-dd8f-48af-9d24-1bf4bafdbe34

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}