Fix: cap scores for sites with single-page sample

[dacharyc]

Fixes #73

Summary

• Apply a 59 (D) score cap when the single-page-sample diagnostic fires, closing the gap that let sites with structurally invalid llms.txt land in the B range from a tiny subset of site-wide signal (single-page-sample diagnostic should apply a score cap #73).
• Cap mirrors the existing no-viable-path precedent in computeCap() — diagnostic-keyed, lowest-cap-wins.
• Update SCORING.md, scoring-reference.md, docs/agent-score-calculation.md, and docs/interaction-diagnostics.md to document the new cap.

Test plan

• npx vitest run test/unit/scoring/score.test.ts passes (added two cases: cap fires at 59; loses to no-viable-path's 39).
• Full suite: npx vitest run (1244 tests pass).
• npx tsc --noEmit clean.
• Manually score the issue's repro site (https://docs.readthedocs.com/platform/stable/) and confirm overall drops from 81 (B) to 59 (D) with cap.checkId === 'single-page-sample'.
• Verify scorecard output renders the (Capped: single-page-sample — …) line.