Fix query9 validator accepting wrong answers and refusals (word-boundary match)

[marc-shade]

Fixes #78.

query_crmarenapro/query9/validate.py used a plain substring test on its fast path:

if expected.upper() in llm_output_clean.upper():   # expected = "MI"

"MI" occurs inside common words in agent prose (determine, minimum, submitted, administrative), so the validator accepted refusals and answers that named a different state.

This switches the fast path to word-boundary matching, which is already the semantics the fallback loop below it uses (re.search(r'\b' + state + r'\b', ...)). One line, no new imports, re is already imported.

Before / after

Same eight inputs, run against the validator before and after the change:

LLM output	before	after
MI	PASS	PASS
The state is MI.	PASS	PASS
the answer is mi (lowercase)	PASS	PASS
I could not determine the answer.	PASS	fail
The answer is CA, at a MInimum.	PASS	fail
The submitted record shows TX.	PASS	fail
CA	fail	fail
The answer is Texas.	fail	fail

All four false positives close; no true positive regresses, including the lowercase and in-prose forms.

Scope

Only this one file. I found it by running all 104 validate.py files against two content-free probes (a refusal containing no numbers, and a string listing many candidate numbers). 85 resisted both; this was the only validator that accepted a refusal. The other 18 accept a numeric shotgun, which looks like the recall-over-precision tradeoff already noted in the paper rather than a defect, so I have not touched them.

Happy to follow up with a small CI test that asserts every validator rejects a refusal, if that would be useful. It would currently need an allowlist for those 18.

[Ruiying-Ma]

Thanks, @marc-shade! We've merged your fix into the validator. Really appreciate it!