fix(incremental): prune edges by source-only to preserve inbound edges from unchanged files

[tirth8205]

Summary

Phase 2 incremental updates previously pruned existing edges whose source or target referenced a removed node. When file F was modified, F was re-analyzed and regenerated its outbound edges — but edges into F from an unchanged file U (e.g. U imports F, U calls F, U tests F) were silently dropped, because U is never re-analyzed and therefore never regenerates them. Every incremental run silently eroded the inbound edges of changed files. In one observed 64-file incremental, 194 inbound edges would have been lost.

This PR addresses Bug 2 of 3 from #366. Bugs 1 (rename detection) and 3 (importMap recovery) are tracked in separate PRs.

The fix

Old rule: prune edges whose source OR target was a removed node.
New rule: keep an existing edge iff its source node survives the node prune (i.e. prune by source only).

Why this is safe

• Inbound file-level edges (U imports F where U is unchanged, F is changed) survive prune. After merge, the fresh batch re-emits file:F, so the target resolves and the edge is dedup'd by (source, target, type, direction). Result: inbound edge preserved.
• Inbound function-level edges into a since-deleted function inside F (function:U:caller → function:F:removedFn) briefly survive prune. The merge script's existing dangling-edge sweep (Step 6 in merge-batch-graphs.py) drops them once function:F:removedFn does not reappear in the fresh batch. Result: stale edges still cleaned up.
• Outbound edges from F are pruned at this step (F is removed → its outbound source is gone). The fresh F batch re-emits them.
• Overlapping edges (an outbound edge from U → F regenerated by U on a future run) hit the merge's dedup on (source, target, type, direction) and pick the heavier weight.

Trade-off: over-keeping. The dangling sweep handles edge cases; a deleted inbound edge from an unchanged source can never be recovered until the next full re-analysis.

Changes

1. understand-anything-plugin/skills/understand/SKILL.md — Phase 2 "Incremental update path", step 2 rewritten with the new rule and a safety-analysis paragraph.
2. understand-anything-plugin/hooks/auto-update-prompt.md — same rule in the auto-update commit hook.
3. understand-anything-plugin/skills/understand/merge-batch-graphs.py — new prune_existing_graph(existing, changed_files) helper so the prune is programmatically testable and agents can call it directly. Does not mutate input. Preserves non-node/edge top-level fields (projectName, frameworks, layers, etc.).
4. tests/skill/understand/test_merge_batch_graphs.py — 8 new regression tests in PruneExistingGraphTests:
   • test_inbound_edges_from_unchanged_sources_survive — the issue's fixture (F1 imports F2, F3 imports F2, only F2 changed → F1→F2 and F3→F2 survive).
   • test_outbound_edges_from_removed_nodes_are_dropped
   • test_function_level_nodes_inside_changed_file_are_removed
   • test_no_changed_files_is_noop
   • test_unrelated_top_level_fields_preserved
   • test_does_not_mutate_input
   • test_merge_dangling_sweep_drops_truly_removed_function_target — end-to-end: inbound edge into a deleted function survives prune, then the dangling sweep drops it after merge.
   • test_merge_preserves_inbound_file_edge_after_prune — end-to-end happy path through prune + merge.

Test plan

• python3 -m unittest tests.skill.understand.test_merge_batch_graphs -v — 77/77 pass (8 new + 69 existing).
• pnpm test — 200/200 pass.
• pnpm lint — clean.
• pnpm --filter @understand-anything/core test — 670/670 pass.

Refs #366

🤖 Generated with Claude Code