databricks-genie skill: add data-grounded Genie Space authoring playbook

[ryanbates99]

What

Adds databricks-skills/databricks-genie/authoring.md — a playbook for authoring a high-quality Genie Space via a fully-curated serialized_space, and wires it into SKILL.md and spaces.md at the natural decision points.

Why

The databricks-genie skill's create guidance stops at the basics: display name, table list, description, and a few sample questions. The manage_genie tool already accepts a complete serialized_space payload, but the skill never teaches the model to author one. As a result, the curation that actually drives Genie answer quality gets skipped:

• column synonyms / value vocabulary
• structured instructions
• certified example question→SQL pairs
• join specs
• reusable measures / filters / expressions
• benchmarks

Spaces created via the skill therefore answer worse than they could, because the model improvises filter values from column names (e.g. status = 'ACTIVE' when the data holds 'active') and leaves the rich layers empty.

What the playbook adds

• Golden rule: ground everything in real data. Inspect with get_table_stats_and_schema, then pull actual distinct values with execute_sql before generating any SQL/instructions — never invent status/category/tier values.
• Per-layer authoring guidance with sensible default counts: table/column descriptions + synonyms, 5 sample questions, GSL-structured text instructions (the five canonical section headers, including the verbatim "Instructions you must follow when providing summaries" string), ~12 example question→SQL pairs with usage_guidance and parameters, 5 measures / 5 filters / 3 expressions, join specs for 2+ tables, 10 benchmarks.
• Exact serialized_space field shapes for every layer (the array-of-lines SQL convention, the --rt=FROM_RELATIONSHIP_TYPE_*-- join tag, parameter type_hint normalization, excluded-column handling).
• API constraints to respect while assembling: at most 1 text_instructions object, ≤25,000 chars per string, ≤30 tables, ≤3.5 MB total, 32-char lowercase-hex unique IDs, and array sorting rules.
• A SQL-validation step — test every example query and benchmark with execute_sql and fix or drop failures before embedding.

This is documentation only and adapts the approach to the kit's existing MCP tool surface (get_table_stats_and_schema, execute_sql, manage_genie). No code changes: the skill installer auto-discovers extra files in the skill directory, and the test manifest asserts no specific file list.

Testing

• Verified the installer (install_genie_code_skills.py) uploads all non-SKILL.md files in a skill directory via the git tree listing, so authoring.md ships automatically.
• Verified the databricks-genie test manifest uses expected_files: [] and does not assert on a file list, so adding a reference file does not break the skill test baseline.
• Cross-checked every documented serialized_space field shape and API constraint against the Genie Conversation API docs (serialized_space schema + validation rules) and the canonical GSL instruction section vocabulary.

This pull request and its description were written by Isaac.