Explore exact combined repository graphs

For the current route/menu map, see the desktop guide.

Purpose and Prerequisites

Use the canonical graph runtime with two or more processed exact repository snapshots in a Brain. Neo4j performs traversal; the companion uses Enola and Tree-sitter/HCL for parsed source evidence. This operation makes no model request and does not resolve remote Git refs.

The worker discovers complete manifest sets and projects cross links automatically. Old publications without parsed witnesses remain queryable but cannot establish remote links from source hints alone.

Procedure

1. Publish the committed repositories through the paired companion. Use a literal Git module source with a full commit and optional subdirectory. The target repository must be registered in the Brain and its snapshot must match that commit. Branch/tag/registry addresses stay unresolved.
2. Create an environment manifest with one exact snapshot per repository. Empty configuration paths select the whole snapshot; otherwise select literal files or directories. Descendants match at a / boundary, without glob expansion.
3. Open Graph → Filters, choose Combined repositories, and select Graph environment and Graph manifest. Graph repositories can narrow the view. Load after the required inputs finish processing.
4. Inspect Exact repository inputs and Unresolved repository links. At most 100 eligible source issues appear with an explicit total. Ambiguous targets, unsupported sources and mismatched commits do not become edges. Identical snapshot sets can share generations across environment manifests.
5. Open Find path, choose endpoints and Find shortest eligible path. Enola's HCL declarations point from symbols to containing directories; both can traverse them in reverse. The path displays recorded direction and marks reverse steps. A terraform_module edge means an exact declared dependency, without claiming initialization, deployment or runtime behavior.
6. Inspect evidence shows the canonical fact and parsed source witness. Rebuild graph queues work for the exact set. Polling does not repeat a mutation. Inspect the existing Processing panel for cancellation and retry.

Verification

The validation mapping records executed proof and remaining closeout. Repeat the synthetic native workflow and regressions with:

./scripts/test-ui.sh tests/graph-combined.spec.ts
./scripts/test-platform.sh
./scripts/validate.sh

The browser fixture publishes committed bytes from two owned local Git repositories, preserves its dirty working copy and verifies a three-hop path without model calls. It does not scan customer repositories.

Failure and Recovery

• Missing input/projection: finish publication/materialization and run the worker. Missing manifest entries never substitute a newer snapshot.
• Unregistered alias: attach an origin only when it identifies the same repository. The link epoch changes and current links rebuild automatically. Conflicting origin ownership cannot silently merge repositories.
• Large view: narrow repositories or configuration paths. The complete view admits 5,000 candidates and 20,000 edges; descriptor loading also has a 64 MiB aggregate budget, narrowed by repository selection.
• Rejected/erased/expired input: canonical eligibility applies before traversal. Rebuilding cannot revive a rejected route. Every required manifest snapshot must remain available even when a narrower display omits its nodes.
• Backend outage/incomplete import: preserve durable work and the privacy journal, recover Neo4j, then retry through the application. Shared endpoints survive removal of another generation. Do not edit Neo4j to repair canonical evidence.

Bounded analytics is available for exact eligible inputs. Desktop graph exploration is also available. Retrieval fusion remains the next slice.