Fix validation script bugs across 13 validators

[anushruthasura]

Summary

Fixes bugs and code quality issues in 13 validation scripts across 7 datasets, discovered during PromptQL benchmark evaluation (8 runs, 2 models).

Window size mismatches (comment says 50, code uses 10 or 20)

• PANCANCER_ATLAS/query1: start+10 → start+50 (matches docstring)
• DEPS_DEV_V1/query1: start+10 → start+50 (matches docstring)
• stockindex/query3: +20 → +50 (matches docstring)

These tight windows cause false negatives when the agent formats output as a table or multi-column layout.

Forbidden-value false positives (stockindex/query1, query2)

• Old: reject if ANY forbidden value appears anywhere — penalizes correct answers that also provide context (e.g., "399001.SZ is most volatile; N225 is least volatile" → FAIL)
• New: reject only if a forbidden value appears BEFORE the correct answer (positional check)

Bare except: → except ValueError: (6 files)

Bare except catches SystemExit, KeyboardInterrupt, etc. Fixed in: GITHUB_REPOS/q1, stockmarket/q1, PANCANCER_ATLAS/q3, yelp/q1, yelp/q2, googlelocal/q3.

Whitespace-stripping match → word-boundary regex (agnews/q1, music_brainz/q2)

• Old: stripped ALL whitespace then used substring matching ("therundown" matched inside "therundownfilm")
• New: case-insensitive word-boundary regex (\bThe Rundown\b)

Test plan

• Wrote 19 unit tests covering all changed validators (pass/fail cases, edge cases)
• All 19 tests pass
• Cross-checked against 216 actual PromptQL benchmark responses (4 runs × 54 queries)
• Verified all 13 files parse and import correctly
• Confirmed result flips are correct (old false positives now properly rejected)

🤖 Generated with Claude Code