The Explorer with change context

The interactive explorer (--explore) has always shown column lineage and the "what breaks if I change this column?" impact panel. It now also surfaces the decision-engine signals — the semantic categorization, the policy verdict, and cross-boundary Metabase reach — when you give it a change context to work from.

Turning on change context

In plain --explore mode (no diff) the explorer looks exactly as it always has. To light up the new surfaces, point it at a changeset the same way the impact command does. These options apply only with --explore:

# Two-manifest diff — the reliable source
dbt-col-lineage --explore \
  --base-manifest base/manifest.json --base-catalog base/catalog.json \
  --policy policy.yml \
  --metabase metabase_lineage.json

Option	What it adds
--base-manifest	Base (target-branch) manifest for a two-manifest diff — enables the change context.
--base-catalog	Base catalog paired with --base-manifest (defaults to a catalog.json next to it).
--git-base <ref>	Alternative: derive the changeset from a git diff against a ref, when no base manifest is available.
--policy <path>	Evaluate the policy so the explorer can show the verdict.
--metabase <path>	Load a metabase_lineage.json snapshot so the explorer surfaces dashboard reach. Consumed offline.

When a change context loads, the server confirms it on startup — e.g. Change context loaded: 6 changed column(s), policy=block. The explorer then shows exactly the signals the CI gate would, computed from the same primitives.

What you'll see

• Policy panel. A dedicated panel for the whole-change verdict, with a pinned banner: block reads as a red stop, warn as amber, allow as a neutral check. Each fired rule expands to show the subject change and the reach it matched, and the selective build/test sets are copy-able (as dbt build --select … chips).
• Semantic badges. Changed columns in the impact panel carry a semantic chip: breaking columns get the amber caution spark; proven-equivalent columns get a de-emphasized neutral check — the good news is deliberately quiet. A column whose semantic is indeterminate is drawn as breaking (fail-safe: the UI never renders "unknown" as "safe").
• Cross-boundary cards. In the affected-exposures list, Metabase dashboards render as a distinct card variant — tagged via Metabase, with a View dashboard link, a meta.tier tag, and a "table-level" caption when the reach is table-grain rather than column-precise.
• Confidence pip. A compact confidence dot sits next to the related-exposures tile on the relationship summary card: neutral + check when the analysis is complete, amber when it's partial — with a tooltip explaining the gap. The full confidence badge and coverage footer in the impact panel are unchanged.

Colour discipline

Amber is rationed to a single caution axis — breaking / warn / partial-confidence all share it and never compete in one view. The one escalation above amber is a policy block (the red stop). This keeps the reading "amber once, plus a red stop when the gate actually blocks".

Status

The panels above (policy panel, semantic chips, cross-boundary exposure cards, confidence pip) are shipped. The graph-canvas marks — amber blast-path edges, a dedicated BI/Metabase boundary band with dashboard nodes, and block/warn rings on the subject node — are a planned polish layer: the API already emits the data (semantic/breaking/boundary on nodes, source on exposures), and the panels already deliver the information at full fidelity.