Incremental update loses cross-file edges from changed→unchanged files (batch-scoped import resolution + replace-on-re-extract) #2406

Version: graphifyy 0.9.19 · Windows 11 · Python 3.14 · TS/TSX corpus (~540 code files)

Symptom

After an incremental update (re-extracting only changed files, then build_merge), the graph silently loses every cross-file edge from a changed file to an unchanged file. In our repo a utility imported by ~every component (v2/src/lib/cn.ts, byte-identical import { cn } from "@/lib/cn" across files) dropped from ~100 edges to 4, and its neighborhood became a disconnected island. ~610 edges were lost in one update.

The failure signature is exact and makes the cause easy to confirm:

Root cause

1. extract() resolves cross-file imports only against the batch it is given (_augment_symbol_resolution_edges(paths, ...)). When the update batch contains only changed files, imports of unchanged files resolve to nothing — those edges are simply never produced.
2. build_merge's replace-on-re-extract then drops the changed file's old (complete) edge set and replaces it with the new (gutted) one.

Each behavior is individually reasonable; combined they guarantee edge loss on every incremental update of a codebase.

Minimal repro

from pathlib import Path
from graphify.extract import extract
from graphify.build import build_merge
# A.ts imports { b } from "./B"
r_full = extract([Path("A.ts"), Path("B.ts")], cache_root=None)   # A→B edge present
r_incr = extract([Path("A.ts")], cache_root=None)                 # A→B edge absent (B not in batch)
# build_merge([r_incr], graph_path=...) replaces A's edges → A→B edge deleted from graph.json

Suggested fix

Since AST extraction is deterministic and cheap (~30 s for 500 files, no LLM), the update path could always run structural extraction over the full corpus and be incremental only for semantic (LLM) extraction. Alternatively, _augment_symbol_resolution_edges could resolve against the existing graph.json node index in addition to the batch.

Secondary finding: dangling edges to never-created node ids

In a full, clean extraction of the same corpus, ~356 edge endpoints reference node ids that no node pass ever creates:

• Barrel re-exports: imports of Button/Surface/Input through v2/src/design/primitives/index.ts emit edges to e.g. ..._primitives_index_button, but no such node exists (the components live in Button.tsx as ..._primitives_button_button).
• Wrapper consts: re-exported consts like const Dialog = DialogPrimitive.Root get edges (..._ui_dialog_dialog) but no node.
• Some of these ids also carry an absolute-path prefix (c_users_<user>_dev_<repo>_v2_src_...) that escaped the relativization/id-remap pass, so they can't even be suffix-matched to real nodes.

Happy to provide the full dangling-id list or run diagnostics against a branch.