feat: add datahub-ml-incident-response skill

[denzasikora-lab]

Adds a catalog investigation skill for observed ML data incidents: freshness breaches, online/offline feature skew, duplicate replay, partial backfills, late events, and deployment/lineage mismatches. The skill verifies DataHub assets and owners, bounds lineage reads, and produces remediation, validation, rollback, and resume conditions from the available evidence.

The skill is registered in the catalog routing guide and exposed through /catalog-ml-incident-response. General dependency questions continue to route to Lineage, and assertion/native incident management to Quality.

Key behavior:

• Read the installed MCP schema before use. The example uses the official server's upstream, max_hops, and max_results parameters. One-hop reads and an explicit total entity budget keep traversal bounded; max_hops=3 is not treated as a strict three-hop limit.
• Accept multiline operational logs as evidence while keeping their contents out of executable shell code. Separate catalog metadata from actual serving-state, watermark, and feature-parity evidence.
• Record observation time, matched feature identity/version/window, graph coverage, incomplete evidence, and verified deployment state in the report template.
• Keep the JSON dry_run report envelope separate from executable DataHub/GitHub API payloads. Requested writes are previewed and handed off only within explicit authorization for the verified target and operation; existing authorization is respected.
• Repair the template's Markdown tables by removing unescaped pipe separators from cell values.

The branch is synchronized with upstream main through a merge, preserving the existing PR history. No release/version files are changed.

Validation:

• All repository pre-commit hooks pass, including Markdown lint, Prettier, YAML/file checks, and Ruff.
• Skill naming/frontmatter/body validation passes with the repository's user-invocable and min-cli-version extensions retained and checked separately.
• Local reference links resolve; the report-envelope JSON parses; each template table has a consistent column count.
• MCP example parameters and the transitive-hop behavior were checked against the official MCP server implementation.

No live DataHub connection or production mutation was used. These are documentation and integration checks; they do not claim a live incident-response evaluation or new model-based behavioral test.