Merge main v4.0 → legacy: self-contained SKILL.md, consistency fixes, human text descriptions

[MADEVAL]

What

Merge all v4.0 improvements from main into legacy:

v4.0 architectural changes

• Self-contained SKILL.md — top-15 per-language data inlined (burned words, AI markers, tone profiles)
• Pre-flight guard — weighted threshold, min text size, RU канцелярит adjustment
• Per-language specificity examples — rung 0→4 for all 9 languages
• Per-tone per-language markers — authentic phrases for 7 tones × 9 languages
• RankWise + MindFluence integration — preservation rules, triple pipeline
• External validation — 3-layer framework (readsight + pymorphy3/spaCy + ZeroGPT)

Consistency fixes (audit)

• Pipeline stage count unified across all files
• «What human text looks like» per-language descriptions restored
• Stage Scores output format — semantic explanations restored
• Missing UK detection marker fixed
• 8 example files: v3 → v4
• validate.sh: version check 3.0 → 4.0
• validate.ps1: UTF-8 encoding fix for verify flags
• PLAN.md: stale references and checkboxes updated
• full-rewrite.md: max consecutive category 3 → 2

New scripts

morph-check.ps1, readability-check.ps1, clause-check.ps1, run-benchmark.sh, run-eval.ps1

---