Fix rendering strategy diagnostics

[dacharyc]

Summary

Two related fixes for rendering-strategy, prompted by Archbee scoring as having a third of its pages "use client-side rendering" when in fact every sampled page was fully server-rendered — they were just brief.

1. Split spa-shell-html-invalid into two distinct diagnostics

The diagnostic previously fired whenever the warn-or-fail proportion exceeded 25%, but its message claimed pages used client-side rendering. That was correct for actual shells but wrong for sparse-but-rendered pages, which is the larger bucket in practice.

• spa-shell-html-invalid now fires only on actual SPA shells (spaShells / total > 0.25) and keeps its strong "client-side rendering invalidates the HTML path" message.
• New sparse-content-html fires on sparse pages (sparseContent / total > 0.25), suppressed when the shell diagnostic already fired. Its message and resolution direct readers to spot-check rather than assume a renderer bug.

2. Loosen the heuristic so legitimately short doc pages pass

detect-rendering previously required visibleTextLength >= 1500 (after chrome stripping) as the wall-of-text fallback. Archbee's short integration explainers (Loom, Figma, Branches & Reviews) sat at 589–1485 chars of real prose and got flagged as sparse. Real shells, by contrast, post-strip to ~0 chars — the gap is much wider than 1500 implied.

Added a new clause to the hasContent disjunction:

(contentHeadings >= 1 && visibleTextLength >= 500)

A heading plus 500+ characters of post-chrome-strip body is well above any actual shell and well below what's needed to admit one. Two new tests cover both sides: a div-soup short page (passes) and an <h1>Loading…</h1> shell with empty body (still fails).

Impact

Archbee re-scored: serverRendered 36→68, sparseContent 32→0, page-size 89→100, overall 78→80 (B).

Doc updates

The split and the proportion math touched several docs. Updated:

• docs/agent-score-calculation.md, docs/checks/page-size.md, docs/improve-your-score.md
• docs/interaction-diagnostics.md (new sparse-content section, updated shell section)
• SCORING.md, scoring-reference.md

Score-cap labels changed from "75%+/50%+ SPA shells" to proportion-based language reflecting the actual formula: (serverRendered + sparseContent × 0.5) / total. SPEC.md needed no changes — it already specifies the three result levels accurately.

Test plan

• All 1,231 tests pass
• Lint and typecheck clean
• Re-ran Archbee end-to-end and confirmed expected score change
• Spot-check at least one other site previously flagged for sparse content to confirm no regression