chore(seo-schema-generator): restore dual-platform code/ + desktop/ structure

The full-merge commit removed 17's code/ and desktop/ folders along with the old
template-fill generator, leaving 17 structurally inconsistent with 16 (and the repo's
documented dual-platform convention). Restore them for the MERGED skill:

- code/CLAUDE.md — Claude Code directive pointing to the root two-mode pipeline
  (no script duplication; root scripts/ stays the single source of truth).
- desktop/SKILL.md + skill.yaml — Claude Desktop directive for the two-mode skill.
- desktop/tools/{firecrawl,perplexity}.md — restored verbatim from git (still used by
  Mode 1 crawl / Mode 2 research).

The old generator's logic stays retired; only the dual-platform folder structure and
relevant tool docs return. 17 now matches 16's layout.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

custom-skills/17-seo-schema-generator/desktop/SKILL.md

@@ -0,0 +1,55 @@
+---
+name: seo-schema-generator
+description: |
+  Generates validation-ready JSON-LD for a site via a claims register — Mode 1 from
+  an existing website, Mode 2 from collected sources for a not-yet-published site.
+  Triggers: generate schema, create JSON-LD, source-to-schema, pre-launch schema,
+  schema from site, claims register, 스키마 생성, 스키마 저작.
+---
+
+# SEO Schema Generator
+
+> Desktop reference. The full, runnable specification (scripts, references, templates,
+> fixtures, stage gates) is the skill-root `SKILL.md` and its bundled directories — this
+> file is the Claude Desktop entry point.
+
+## Purpose
+
+Author JSON-LD for a site whether or not its pages exist yet. Both scenarios route
+through one pivot — a **claims register** (provenance-tracked, conflict-resolved facts) —
+then generate pruned drafts that hand off to **seo-schema-validator** (generate → validate).
+
+## Two modes, one pipeline
+
+| | Mode 1 — existing site | Mode 2 — collected sources |
+|---|---|---|
+| Source of truth | the live pages | scattered sources (DART, Wikidata, brochures) |
+| Seed the register with | extract from pages | manual research |
+| Hard part | extraction & mapping | authority hierarchy + entity reconciliation |
+
+Everything after the claims register (build drafts → prune unfilled slots → validate)
+is identical. **Only CONFIRMED, non-conflicting claims become schema;** unfilled template
+slots are deleted, never shipped as placeholders.
+
+## MCP Tool Usage
+
+```
+mcp__firecrawl__scrape / crawl : Mode 1 — pull existing pages to extract facts
+mcp__perplexity__search        : Mode 2 — discover & cross-check authoritative sources
+```
+
+## Workflow
+
+1. Lock the entity→type map (scope first).
+2. Seed the claims register (Mode 1: extract from pages · Mode 2: research → register).
+3. Reconcile to CONFIRMED; clear conflicts.
+4. Build drafts from type templates (placeholders pruned).
+5. Validate with seo-schema-validator — gate = zero P0.
+6. Fix P0, re-validate, then client review against the report (not raw JSON).
+
+## Notes
+
+- Mode 1 inference (title/OpenGraph) is seeded PENDING and never auto-ships; existing
+  JSON-LD is seeded CONFIRMED. If a site already has good JSON-LD, audit it with
+  seo-schema-validator (Mode B) instead of regenerating.
+- Authoritative rich-result eligibility still needs Google's online test on a sample.