test(experts): close the mutation gaps that survived the full suite

[kevincostner17]

What this closes

A fresh mutation measurement of semantic/experts.py, using a rebuilt harness validated against a known baseline, gave 267 sites, 64.4% subset kill rate, 95 survivors. A subset survivor is only a candidate, so all 95 were re-run against the full suite:

outcome	count
killed by the full suite (all number/month table entries — pinned by #504)	47
real gaps — observable, now each killed only by its dedicated test	13
equivalent — exhaustive executable proof	10
in-range confidence/evidence literals — deliberately not pinned	22
audit-note sampling cap — a separate finding, not pinned	1

The survivors don't sum to 48 because some tests cover two mutants of the same line.

The real gaps

Each was shown to be observable before any test was written: the original and the mutant were run on the same input and gave different outputs.

• "(1,200" (unbalanced) read as the accounting negative −1.2, flagged unambiguous
• "($1.00)" read as 0.0, a one-unit debit erased; "($0.00)" read as −0.0 or 0.05
• malformed grouping accepted: "$1,2,345" → 12345, "$,123" → 123
• a bare "hundred" defaulting to 200
• the leading-zero identifier veto checking the wrong character: "05" and "0123" lost protection and "10" gained it. Every existing test used "007", which has a zero in both places.
• valid ISO date strings rewritten to timestamps beside a value that can't be resolved, contrary to the code's own "no representation churn" rule
• the partial-date review guess becoming year 0003
• resolver and expert base confidences above 1.0, hidden by the downstream 0.999 clamp. These are pinned as a [0, 1] invariant, not as literals.

How the tests were checked

• Kills: each targeted mutant is killed only by its own dedicated test.
• No false kills: every mutant classified equivalent leaves the whole file passing. This check caught an early draft that used source-text tripwires (inspect.getsource). Those failed on every mutant, because the harness writes mutants through ast.unparse, which reformats the file, so they would have silently inflated every future mutation score. They were removed.
• Equivalence proofs check a finite domain exhaustively and state their premise in prose. None relies on "tried some inputs, found no difference".
• Confidence literals aren't pinned: pinning a literal only detects change. A decision/risk-tier fingerprint over a 10-expert corpus found no change for any of them. That corpus detects 1 of 5 known behaviour-changing mutants, so this is recorded as weak evidence, not as equivalence.

Verification

• 53 tests, all passing on py3.12 / pandas 2.3.3 and py3.9 / pandas 1.5.3
• 334 passed under each of 5 pytest-randomly seeds, alongside the neighbouring experts tests
• Full suite, 0 failed: py3.12 7592 passed, py3.9 7560 passed
• ruff check clean
• Test-only change

[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: ca7ab4e4-d39c-44fd-b98e-8e2248259393

---

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.}