Add vertical intelligence flywheel skills and artifacts

[danielfoch]

Summary

Adds a small Vertical Intelligence Flywheel layer to GStack: skills and schemas for compressing repeated context, capturing human-approved runs as redacted traces, reviewing data-heavy work, and exporting training-ready examples for optional future open-weight SLM/adapters.

Why

A lot of OpenClaw/GStack usage happens inside vertical businesses: real estate, legal ops, local services, healthcare admin, insurance, finance ops, etc. The repeated work is valuable. If users can capture approved runs, compress context, create evals, and export clean traces, they can gradually build private vertical intelligence without adding telemetry or training anything inside GStack.

The goal is:

approved run → compressed context → eval case → training-ready trace → optional future SLM/adapter

What changed

• Added /context-compress
• Added /trace-to-train
• Added /data-review
• Added schemas for traces, context cards, eval cases, training examples, and flywheel metrics
• Added docs for the Vertical Intelligence Flywheel
• Added a sanitized realtor/OpenClaw example as one worked vertical

Non-goals

• No model training
• No telemetry
• No external API calls
• No raw PII storage
• No hard-coded model provider
• Not real-estate-only

Privacy model

Local-first, opt-in, redacted by default. The skills help users decide whether a run is retrieval-ready, eval-ready, or trainable. Anything client-facing, compliance-sensitive, or PII-bearing requires human review.

Model stance

Model-agnostic. Gemma or another open-weight model can be a downstream target later, but this PR only creates the context/eval/trace foundation.

Example

A realtor OpenClaw provider can capture approved lead-intake, listing-copy QA, CRM follow-up, and transaction-checklist runs. Over time those traces can become compact context cards, evals, and training examples for cheaper vertical inference.

Test plan

• Ran bun run gen:skill-docs --host all
• Ran bun run skill:check
• Validated JSON schemas and redacted trace example with jq empty
• Validated realtor eval YAML parses
• Ran git diff --check
• Ran focused Bun tests

Focused Bun tests were run, but the suite reports an existing large tracked fixture failure on browse/test/fixtures/security-bench-haiku-responses.json, which is outside this PR.

Full bun test was also attempted, but this local workspace has unrelated failures from the repo path containing spaces, missing git author config in temporary test repos, and existing browser test assertions outside the touched files.