Add GeoSkills v0.5 data QA and classification foundation

[gronbow]

Summary

• add count-only, privacy-safe geochemical data QA for sample identifiers, missing/BDL states, invalid cells, nonpositive values, and opt-in major-oxide totals
• add deterministic same-unit derived ratios without arbitrary formula evaluation or implicit mixed-unit conversion
• add a versioned generic classification-model foundation with provenance, scientific-review gates, SHA-256 pinning, log-axis support, geometry validation, and boundary/overlap review
• integrate QA policies, derived definitions, summaries, confirmations, and plan invalidation into the existing recipe → plan → run → report workflow
• update the Skill instructions, example recipe, user documentation, changelog, and interface metadata

Scientific and privacy boundaries

• duplicate sample identifiers remain a blocking error
• major-oxide total thresholds and composition basis must be explicitly supplied and reviewed
• no new formal classification diagram or literature boundary is registered in this phase
• shareable plans and reports contain counts and definitions, not sample identifiers, exact source values, or absolute paths
• local/private data and acceptance outputs are excluded from this PR

Validation

• Skill structure: quick_validate.py → Skill is valid!
• automated tests: 217 passed
• synthetic end-to-end major-element workflow: plan ready, run ready, 14 output files
• scoped shareable-artifact privacy scan: 7 files checked, 0 sample-ID or absolute-path leaks, 0 sensitive source-data CSV files
• existing REE, spider, Harker, TAS, synthetic, and published-data regressions remain passing

Release status

The candidate review is complete and the user has separately approved merge, the v0.5.0 tag, and a formal GitHub Release. The PR now contains the final 0.5.0 version metadata and dated changelog. Merge only after the updated cross-platform CI matrix passes; create the tag and Release from the verified merge commit.