fix(fixtures): scorer regex accepts digits + first validated baselines (6 scenarios, avg 99/100)

[ptengelmann]

Summary

Real-API bench run across the five new scenarios from #15, plus a regex bug fix the run surfaced, plus the first validated marketing claims that survive the truth-check.

Bench results

All six scenarios scored EXCELLENT (90+):

Scenario	Score	Signal	Code	Unexpected
freehold-enforcement-and-undisclosed-occupier	97.0	4/4	5/5	1
freehold-clean-with-satisfied-restriction	97.0	3/3	1/1	1
freehold-resolved-boundary-dispute	100.0	2/2	1/1	0
freehold-missing-building-regs-cert	100.0	3/3	2/2	0
freehold-disclosure-inconsistency-flooding	100.0	4/4	4/4	0
leasehold-short-lease-and-escalation	100.0	5/5	5/5	0

Average: 99.0 / 100. 21/21 signals across all six. 18/18 codes across all six. 0/8 over-flags across the three scenarios that have antipatterns.

The two scenarios at 97 lose 3 points each to a single "unexpected" code emission: PLANNING_BUILDING_REGS_MISSING (defensible inference from the enforcement notice) and TITLE_RESTRICTION_CONSENT (model emitted the code for the Form A restriction but correctly placed it at LOW severity, so the code emission is observation not over-flagging).

Scorer regex bug discovered + fixed

The leasehold scenario initially reported 4/5 codes hit. Investigation showed the pipeline emitted the right code ([CODE: LEASE_SECTION_20_MAJOR_WORKS]) in the right place. The scorer's extractEmittedIssueCodes() used [A-Z_]+ which silently dropped any code containing a digit - so LEASE_SECTION_20_MAJOR_WORKS (with "20") was invisible. Fixed by adding 0-9 to the character class. Leasehold scenario re-scored at 100/100 after the fix.

What this changes

• packages/fixtures/src/score.ts: regex character class [A-Z_]+ -> [A-Z0-9_]+ in extractEmittedIssueCodes. With a comment explaining why so the same bug doesn't get reintroduced.
• packages/fixtures/BASELINES.md: six validated baselines (refreshed enforcement scenario + five new ones) with metrics, regression rules, and a summary table.
• apps/marketing/src/sections/trust.tsx: "Severity calibrated" card body now states "Validated against six adversarial conveyancing scenarios with an average Matter Quality Score of 99 out of 100" (both numbers grep-verifiable in BASELINES.md).
• apps/marketing/src/sections/faq.tsx: wrapper answer now references the six validated scenarios with the concrete average score and the "every planted issue detected / zero over-flagging" claim.

Test plan

• pnpm typecheck - 9/9 workspaces pass
• pnpm check - 0 errors, 3 pre-existing warnings
• pnpm build - 3/3 tasks pass
• Real-API bench run on all five new scenarios + re-score of existing scenario
• All six scenarios verdict EXCELLENT
• No new em-dashes; no §; UK English throughout

Cost of this PR

~$4.50 of Anthropic API calls (5 fresh scenario benches at ~$0.90 each).

Honest gaps in the corpus

• 3 legacy scenarios (leasehold-flat-with-issues, freehold-house-clean, freehold-house-edge-cases) have NOT been re-benched in this session. They are in the scenarios registry but not in BASELINES.md. Status: TODO for a follow-up run.
• 5 of 21 taxonomy codes are still not exercised by any scenario (TITLE_OVERRIDING_INTERESTS, TITLE_DEFECTIVE_INDEMNITY_REQUIRED, LEASE_ALTERATIONS_CONSENT, SEARCH_CONTAMINATED_LAND, OCCUPIER_TENANCY_UNDISCLOSED, MORTGAGE_OFFER_EXPIRY_RISK). Growth should follow pilot patterns, not speculation.

Brand and liability check

• All marketing claims verifiable in BASELINES.md
• No "the product is proven" claim - the language is "validated against six scenarios", not "proven across all conveyancing"
• UK English, no emoji, no hype

Moat level

L5 / L7 - the proprietary bench corpus crosses from 1 measured to 6 measured, and the marketing now has measured numbers behind it for the first time. Each baseline metric is a regression guard for future prompt or model changes.