Backport cluster graphs and multigraph mode (T15)

[TimothyNguyen]

Summary

• Cluster graphs. A cluster.json spec names member repos and typed cross-repo links (api_call, shared_resource, mirrored_file, depends_on, references). kb-core cluster init/add/remove/locate/build/check/status manages it. A build composes the members into one directed graph plus a manifest and report, and writes a cluster-ref.json marker into each member so --cluster [NAME] resolves the composed graph from inside any of them — wired into query, path, explain and affected.
• Multigraph mode. Parallel edges between a node pair survive build, analysis, query and export, each keyed by a content-derived stable_edge_key instead of NetworkX's positional key, so a rebuild does not renumber every edge when the extraction order shifts.
• Both opt-in. The default single-repo path is unchanged. The tests that pin that default (test_default_build_stays_simple, test_build_merge_defaults_to_simple_without_a_stored_flag) ship in this PR rather than as an afterthought.

kb-core is a stripped fork of graphify and never received upstream PR Graphify-Labs/graphify#2134. The scaffolding was already present but dormant — require_multigraph_capabilities() had no call sites and link_shared_type_declarations() was never invoked. Without a composed graph.json there is nothing for kb-core-ui to render across repos.

New modules: kb_core/cluster_graph.py, kb_core/cluster_cli.py, kb_core/cluster_ref.py. Touched: build, global_graph, cluster, analyze, affected, export, watch, serve, manifest_ingest, cli, __main__.

Test plan

• pytest tests/test_cluster_spec.py — 41 passed
• pytest tests/test_cluster_links.py — 20 passed
• pytest tests/test_cluster_build.py — 20 passed
• pytest tests/test_cluster_refs.py — 28 passed
• pytest tests/test_cluster_cli.py — 33 passed
• pytest tests/test_multigraph_build.py — 21 passed
• Manual end-to-end: two extracted repos, cluster init, add x2, one declared api_call, check exit 0, build producing graph + manifest + report + per-member markers, then query --cluster and affected --cluster crossing the repo boundary via calls_api. graph.json byte-stable across a --force rebuild.
• 13-file CLI/hook regression name-diffed against the baseline worktree at 6c3f0e2 — identical failure sets (72 pre-existing on both sides).
• Full-suite name-diff against the baseline worktree. Recorded baseline: 154 failed, 4404 passed, 173 skipped. Counts alone are not the gate; the failure names must match.
• cd kb-core-ui && pytest python/tests -q for shared-import breakage.

The two unchecked items and the reasoning behind them are written up in spec/T15-CODEX.md, along with three known non-bugs so they are not re-investigated.

Generated with Claude Code
via Happy