Lineage engine

services.analyze loads inventory metadata and a hash/config/path keyed parse cache. It parses only changed input, then materializes a complete workspace graph in one transaction. A failed materialization rolls back facts and leaves a failed run record; original source is unaffected. Selected analysis reuses caches for unselected files; changed unselected files stay pending.

Operations have explicit input/output memberships. Asset-to-asset edges reference an operation, statement, file and analysis run. Store.edges joins evidence code and exact line ranges. Original bytes are available through Store.original; the UI shows original source separately from extracted statement snippets. Step numbers are dependency ordering, not execution order.

Column transformations are an additive contract on Operation.derivations, persisted by transformations.materialize_columns in the same analysis transaction. column_catalog and trace_columns provide shared, indexed asset/column queries. The traversal keeps bounded visited/scheduled column/effect states, not all-workspace paths. CTE scopes, SAS assignment versions, branch evidence, value-versus-row effects and cycle/limit boundaries survive each step. Multiple operation writers stop upstream chaining with an ambiguity marker. Sources are loaded lazily and snippets/expressions are capped at 12,000 characters. Parsing errors for previously analyzed sources abort refresh and retain committed facts. See column transformation coverage and query semantics.

Identifiers preserve display spelling. Lookup defaults to case-insensitive canonical names; case-sensitive matching can be configured. An unqualified reference links only when exactly one qualified candidate exists; that link is inferred, with an alias record and lower confidence. Ambiguity is a diagnostic and remains a separate node. CSV linking requires an explicit path.

LineageGraph groups parallel evidence on a directed NetworkX graph. Breadth-first traversal has depth/node budgets. Strongly connected components report cycles without enumerating every cycle. Temporary contraction combines evidence and intermediate names with a fanout budget. Endpoints and self-cyclic temporary nodes are retained to avoid losing information.

path_report traverses the selected full/simplified graph with a bounded pending queue (1–500 paths, 1–30 edges deep). Results mark root/terminal/depth/cycle boundaries and global truncation; they do not pretend a depth cutoff is a true root. Ordered steps retain immutable source hash, statement/file IDs, original line ranges and intermediate evidence references. CSV flattens paths/steps/evidence; JSON retains the path hierarchy. Full source bodies are not duplicated for every path. root_paths remains available for existing callers.

ui.lineage_tree.render_tree tracks native expander state and renders only open branches in the active direction tab. Expand/collapse controls change a widget generation counter; only control state, never a rendered tree or source content, lives in session state.