Analyze exact eligible graph inputs

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

Purpose and Prerequisites

Run PageRank centrality, Leiden communities or connected components through the existing local Neo4j/GDS service. The API and worker must be running, and the selected repository, combined manifest or knowledge generation must be ready. Start the local application with ./scripts/dev.sh and open http://127.0.0.1:8787. Credentials remain in the ignored .env.

These reports describe recorded topology or evidence relationships. They do not accept claims, establish operational impact or call an LLM. Normal autonomous memory maintenance continues under the Brain's standing policy.

Procedure

1. Open Graph → Filters in the Brain. Select its graph kind, exact repository snapshot or environment manifest, evidence mode and scope. Wait for current projection processing to finish, including automatic rebuilds after a correction.
2. Open Insights and choose at least one Graph relation explicitly. Structural relations and knowledge provenance form separate analytical inputs. Combined repository reports may include validated terraform_module links.
3. Select Analysis and Analysis direction. PageRank supports outgoing, incoming and both; communities and connected components use both. Every recorded edge has unit weight, including parallel edges and self loops. Isolated eligible vertices are retained.
4. Choose Queue analysis. Writers can queue and cancel; readers can open eligible reports. The shared heavy worker processes one job at a time. Watch progress and cleanup state, then Open analytical report.
5. Expand Exact analytical inputs and parameters to inspect the actual GDS version, recipe, generation, exact inputs and coverage. Inspect analytical evidence opens the retained canonical evidence for a result. Group numbers belong only to that report.
6. Refresh report reads current state without queuing another calculation. If inputs change, old numerical rows disappear. After current graph processing finishes, queue a fresh report. A newer ready generation can invalidate a report even when the displayed eligible vertices look unchanged.

Reports admit at most 10,000 canonical candidates, 50,000 selected relationships and a combined 64 MiB native memory estimate. The read page contains at most 100 results and the metadata page at most 20 reports. Oversized inputs are refused; the application does not publish a truncated aggregate. Fixed PageRank iterations are not a convergence guarantee. Partial extraction remains explicitly partial.

Verification

The contract defines the complete behavior and the evidence mapping records executed results and limitations. With repository-owned services running:

set -a
source .env
set +a
cargo test -p recollect-server --test platform graph_analytics_ -- --ignored --test-threads=1
./scripts/test-ui.sh tests/graph-analytics.spec.ts
./scripts/validate.sh

These checks use isolated databases and synthetic native graphs. They cover complete input invalidation, ordinary-reader concurrency, delayed native creation, interrupted attempts, replaced leases and retained-journal recovery from an actual older PostgreSQL copy. The termination fixture uses a bounded synthetic stress recipe to observe cancellation; it does not change product parameters.

Failure and Recovery

• Empty or unavailable input: finish processing and choose a ready selection with retained evidence. Strict accepted mode excludes proposed claims.
• Stale or removed report: queue a new report after current graph processing. Old scores cannot be salvaged by hiding the removed contributor.
• Resource refusal: narrow scope or relationships. Native estimates are an admission check, not a guarantee against all memory pressure.
• Cancellation or cleanup pending: cancellation is requested immediately; Neo4j can take time to retire its transaction. Automatic maintenance retries exact scratch cleanup. New allocation and privacy completion wait for verified absence. Do not interpret a cancelled job as proof of physical removal.
• Missing or conflicting ownership: restore the matching private journal. Keep .data/erasure-journal, including its analytics-<installation> directory, together with canonical database and artifact backups. Do not fabricate an installation marker or remove an unexpired attempt record to bypass recovery.
• Older database restore: keep the newer retained journal, stop API/worker serving and run recollect-server privacy-reconcile with the operator database configuration loaded, following the privacy runbook. Reconciliation uses journal ownership even when the database predates an analytical attempt. Preserve uncertain state for retry; do not delete native graphs by prefix or touch another installation's catalog entries.

Each attempt has a 120-second deadline and separately bounded compute/control queries. Closed ownership remains retained through the creation deadline so delayed requests cannot resurrect untracked scratch state. Reports are capped at 100 per Brain with 20 pending; terminal reports are pruned after seven days or when capacity is needed. Active work and unresolved scratch are not evicted.