feat: add evidence-backed incident response skill

[equinoxaifinance-rgb]

Summary

Adds a dedicated datahub-incident-response skill for the symptom-to-containment loop:

• anchors a consumer-visible symptom to an exact DataHub entity;
• gathers bounded upstream lineage and metadata evidence;
• separates observations, hypotheses, and confirmed production causes;
• identifies owners and the smallest confirming check;
• maps a complete downstream blast radius;
• proposes exact warning-tag writes and waits for explicit approval;
• verifies every mutation and supports provenance-safe rollback.

The contribution also adds routing, a Claude Code command, focused references, and three adversarial evaluation fixtures.

Why this is separate from lineage and quality

datahub-lineage answers dependency questions. datahub-quality manages assertions and native DataHub incidents. This skill begins with a live consumer symptom and coordinates evidence-backed diagnosis plus temporary, reversible catalog warnings. It does not claim that tags repair a pipeline or that metadata alone proves a production root cause.

Safety and correctness

• Catalog descriptions, SQL, links, and custom properties are treated as untrusted evidence.
• Search results are resolved to fetched URNs before any proposed write.
• Lineage is bounded and pagination/truncation must be resolved before graph-wide containment.
• Plans above 50 targets require a fully enumerated scope and specific approval.
• A pre-plan instruction such as “contain it now” is not approval of undisclosed targets.
• Mutation success is not persistence proof; every target is re-fetched.
• Partial failure stops later batches and reports verified, pre-existing, failed, unverified, and unattempted pairs separately.
• Rollback removes only pairs verified as newly added by that run.
• When root cause is unresolved, the skill may offer a one-entity symptom-anchor warning without speculatively quarantining upstream assets.

Validation

• Repository pre-commit hooks pass on all 10 changed files.
• The tool contract was checked against mcp-server-datahub==0.6.0: get_entities, get_lineage, get_lineage_paths_between, add_tags, and remove_tags.
• Three evaluation fixtures cover malicious catalog instructions / insufficient evidence, a truncated 140-asset blast radius, and partial write failure / rollback.
• Two independent forward tests were rerun after fixes; both returned READY with no blocker.
• The reference implementation in Lineage Detective passes its complete 214-test Python suite and a fresh deployed end-to-end run through live DataHub evidence, model-grounded diagnosis, mutation readback, sandbox verification, exact-byte apply, handoff generation, and hash-verified restore.

Provenance and overlap

This workflow was publicly shipped in Lineage Detective on July 19, before PRs #56 and #61 opened. I am not alleging copying; the implementations independently converge on a real need. This version is intentionally differentiated by OSS-compatible tag containment, prompt-injection boundaries, bounded target enumeration, exact approval semantics, read-after-write verification, partial-failure accounting, and provenance-safe rollback.

Scope

This PR does not create tag vocabulary automatically and does not administer Cloud-only native incidents. It remains useful in read-only mode when mutation tools are unavailable.