Add DataHub-grounded TeaQL modeling skill

[philipgreat]

Problem

The current build-teaql-app workflow is optimized for greenfield modeling from natural-language requirements. When an organization already maintains schemas, classifications, glossary terms, ownership, or lineage in DataHub, the agent should ground TeaQL modeling in that existing context instead of reconstructing the domain from prompts alone.

The hackathon prototype in datahub-hackson-2026 demonstrates the core flow, but its reusable skill currently mixes DataHub acquisition with duplicated TeaQL evaluation and generation instructions, and it does not consistently distinguish source facts from modeling inference.

Proposed outcome

Add a focused, publishable skill to this repository that:

• queries DataHub metadata through an available MCP integration;
• saves the retrieved context as evidence before modeling;
• treats existing context as authoritative for business facts;
• distinguishes datahub-fact, agent-inference, teaql-framework, and unresolved decisions;
• creates a complete local KSML model and a traceable grounding ledger;
• preserves the repository model-first rule: save the complete KSML target before TeaQL evaluation;
• hands evaluation, repair, generation, implementation, and verification to the existing build-teaql-app workflow instead of duplicating toolchain rules;
• never claims runtime enforcement from metadata propagation alone.

Acceptance criteria

• Add a new focused skill under skills/ with valid SKILL.md and agents/openai.yaml.
• Define evidence precedence for structured tags, glossary/classification, lineage/foreign keys, descriptions, and agent inference.
• Define the grounding ledger format and require evidence locators for DataHub-derived facts.
• Define failure behavior when required metadata or relationship evidence is missing.
• Compose cleanly with build-teaql-app and avoid copying versioned TeaQL commands.
• Update the repository entry documentation to expose the new skill.
• Validate the skill package and run repository checks.
• Forward-test the skill against the payment example without treating expected conclusions as hidden context.

Source prototype

The initial design is informed by the local datahub-hackson-2026 payment example: captured MCP entity metadata, generated KSML, decision notes, and context-to-code mapping. The production skill should preserve the useful evidence chain while removing hackathon-specific paths and unsupported claims.

[7in10]

这是来自QQ邮箱的假期自动回复邮件。   您好，无法亲自回复您的邮件。我将尽快给您回复。

[philipgreat]

Implementation is available on feature/datahub-grounded-modeling at commit ba0af0a.

Completed:

• added model-teaql-from-datahub with valid skill metadata;
• added a grounding contract and decision ledger covering datahub-fact, agent-inference, teaql-framework, and unresolved;
• defined structured-metadata precedence, relationship evidence boundaries, sanitization, missing-evidence behavior, coverage checks, and the existing-model delta workflow;
• composed with build-teaql-app instead of duplicating versioned TeaQL commands;
• updated AGENTS.md and README.md with routing, installation, and invocation guidance.

Verification:

• quick_validate.py: passed;
• git diff --check: passed;
• local skill and README links: passed;
• no TODO placeholders or copied versioned TeaQL commands: found none;
• credential-pattern scan over changed content: found none.

Forward testing:

• ran two independent grounding passes against only the payment user request and captured DataHub context;
• the model preserved payment_account as a scalar because no FK/documented join proved a relationship;
• free-text masking was labeled as inference, not a structured tag or runtime proof;
• unrelated owner information was sanitized;
• after iteration, source descriptions were no longer copied into KSML comments;
• all four selected payment fields were covered; missing PII tags, relationship evidence, and DECIMAL precision/scale were reported as evidence gaps;
• TeaQL evaluation and generation were intentionally not run during the grounding-only forward tests.

The pre-existing untracked docs/ directory was not staged or modified.