fix: purge dataflow rows keyed by call_edge_id before edge deletion

[carlos-alm]

Summary

• purgeFilesData / purgeFileData (the WASM/JS incremental file-deletion path) had a gap: dataflow rows whose call_edge_id referenced a calls edge touching the deleted file — but whose own source_id/target_id belonged to other files (inter-procedural stitching) — survived the existing source/target and vertex-based deletions.
• With better-sqlite3's default PRAGMA foreign_keys = ON, the subsequent DELETE FROM edges step then failed silently because those surviving dataflow rows held a FK reference on call_edge_id. Stale nodes remained, causing a node count mismatch after file deletion in incremental builds.
• Fix: add a dataflowByCallEdge purge step in preparePurgeStmts (src/db/repository/build-stmts.ts) that removes any dataflow row whose call_edge_id references an edge touching the deleted file. The step runs just before the edges delete so the FK constraint is always satisfied.
• The companion fix for the native orchestrator path (disabling PRAGMA foreign_keys before buildGraph) was already merged as prerequisite fix: issue-1174 incremental parity test fails — imports edge count mismatch (11 vs 6) #1644.
• A pre-existing role-classification parity failure in build-parity.test.ts was discovered and filed as issue fix: build-parity test fails — native classifies main/square as dead-unresolved but WASM says leaf #1661 (unrelated to this change).

Test plan

• npx vitest run tests/integration/incremental-edge-parity.test.ts — all 10 tests pass, including file deletion > node count matches
• npm test — 3271 passed, 1 pre-existing failure (build-parity.test.ts > produces identical roles, tracked as fix: build-parity test fails — native classifies main/square as dead-unresolved but WASM says leaf #1661)
• npm run lint — no issues

Closes #1645

[greptile-apps]

Greptile Summary

This PR fixes a FK-constraint gap in the WASM/JS incremental file-deletion path: dataflow rows whose call_edge_id referenced a calls edge touching the deleted file — but whose own source_id/target_id lived in other files — survived the existing purge steps, silently blocking the subsequent DELETE FROM edges and leaving stale nodes. The fix adds a dataflowByCallEdge purge statement that runs just before the edges delete. The PR also removes dynamic-dispatch detection from all Rust and JS extractors and tightens the native-orchestrator FK restore with a try/finally.

• src/db/repository/build-stmts.ts: Adds dataflowByCallEdge purge step using a named-param @f query that deletes dataflow rows via their call_edge_id → edges → nodes(file) chain; runs before edges delete to satisfy the FK constraint.
• Rust + JS extractors (c, cpp, go, php, python, ruby, elixir, lua, swift, csharp): Removes all language-specific dynamic-dispatch detection patterns, reducing each call handler to straightforward name extraction.
• src/presentation/queries-cli/overview.ts: Simplifies printRoles but loses the hasDeadSubRoles deduplication guard, causing double-counting of dead-role totals (see inline comment).

Confidence Score: 4/5

The core database fix is correct and well-ordered, but the simplified printRoles function introduces a definite display regression in the CLI overview output.

countRoles() in module-map.ts always adds a synthetic dead aggregate key on top of the individual dead-callable/dead-leaf sub-role keys. The removed hasDeadSubRoles guard was the only thing preventing printRoles from summing all of them together. On any database that has dead-role nodes, the classified symbols total shown by codegraph overview will be inflated.

src/presentation/queries-cli/overview.ts - printRoles should either restore the hasDeadSubRoles exclusion or countRoles should stop injecting the synthetic dead aggregate key when sub-role keys are present.

Important Files Changed

Filename	Overview
src/db/repository/build-stmts.ts	Core fix: adds dataflowByCallEdge purge step with correct SQL and named-param @f binding, placed before edges delete to satisfy the FK constraint on dataflow.call_edge_id.
src/presentation/queries-cli/overview.ts	Simplified printRoles removes the hasDeadSubRoles deduplication guard; since countRoles() always injects a synthetic dead aggregate key alongside individual dead-* sub-role keys, the total is now double-counted.
src/domain/graph/builder/stages/native-orchestrator.ts	Wraps buildGraph() in try/finally to guarantee PRAGMA foreign_keys = ON is restored even on throw; addresses prior review threads.
src/domain/analysis/roles.ts	Removes AND confidence = 0 filter from dynamicCallsData; consistent with the extractor simplification that drops dynamic-dispatch emission entirely.
src/extractors/c.ts	Removes dlsym and (*fp)() dynamic dispatch detection; consolidated to a single Call object built for both field_expression and default cases.
crates/codegraph-core/src/extractors/c.rs	Removes dlsym/function-pointer dynamic dispatch detection; simplifies to plain call recording matching the JS extractor change.
src/features/dataflow.ts	Minor JSDoc wording fix; no logic change.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[purgeFileData / purgeFilesData] --> B[embeddings]
    B --> C[cfgEdges / cfgBlocks]
    C --> D["dataflow (source_id / target_id)"]
    D --> E["dataflowByVertex (source_vertex / target_vertex)"]
    E --> F["dataflowByCallEdge NEW - call_edge_id -> edges touching file"]
    F --> G[dataflowSummary / dataflowVertices]
    G --> H[complexity / nodeMetrics / astNodes]
    H --> I["edges - FK constraint satisfied"]
    I --> J[nodes]
    J --> K[fileHashes]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[purgeFileData / purgeFilesData] --> B[embeddings]
    B --> C[cfgEdges / cfgBlocks]
    C --> D["dataflow (source_id / target_id)"]
    D --> E["dataflowByVertex (source_vertex / target_vertex)"]
    E --> F["dataflowByCallEdge NEW - call_edge_id -> edges touching file"]
    F --> G[dataflowSummary / dataflowVertices]
    G --> H[complexity / nodeMetrics / astNodes]
    H --> I["edges - FK constraint satisfied"]
    I --> J[nodes]
    J --> K[fileHashes]

Loading

_{Reviews (6): Last reviewed commit: "fix: resolve merge conflicts with main" | Re-trigger Greptile}

[greptile-apps]

PRAGMA foreign_keys = OFF applies to the entire ctx.nativeDb connection for the rest of its lifetime, not just the buildGraph() call. If any JS-side code later writes to ctx.nativeDb (e.g., post-build gap-repair steps that share this connection), those writes will also run without FK enforcement. The comment "restored automatically when this connection is closed" is accurate for SQLite's session-scoped PRAGMAs, but could be misread as an explicit teardown. Worth verifying that nothing after buildGraph() writes to nativeDb expecting FK protection, or adding ctx.nativeDb.exec('PRAGMA foreign_keys = ON') right after buildGraph() returns to narrow the scope.

[carlos-alm]

Fixed — added ctx.nativeDb.exec('PRAGMA foreign_keys = ON') immediately after buildGraph() returns (before JSON.parse), so FK enforcement is restored as soon as the Rust-side purge+build is done. The comment was also updated to reflect this explicit restore rather than the previous wording that implied it was only restored on connection close.

[github-actions]

Codegraph Impact Analysis

17 functions changed → 37 callers affected across 24 files

• PurgeStmts.dataflowByCallEdge in src/db/repository/build-stmts.ts:9 (0 transitive callers)
• preparePurgeStmts in src/db/repository/build-stmts.ts:29 (8 transitive callers)
• runPurge in src/db/repository/build-stmts.ts:99 (8 transitive callers)
• dynamicCallsData in src/domain/analysis/roles.ts:15 (2 transitive callers)
• tryNativeOrchestrator in src/domain/graph/builder/stages/native-orchestrator.ts:2089 (5 transitive callers)
• handleCCallExpression in src/extractors/c.ts:145 (2 transitive callers)
• handleCppCallExpression in src/extractors/cpp.ts:250 (2 transitive callers)
• handleCsInvocationExpr in src/extractors/csharp.ts:230 (2 transitive callers)
• handleElixirCall in src/extractors/elixir.ts:53 (2 transitive callers)
• handleGoCallExpr in src/extractors/go.ts:229 (2 transitive callers)
• handleLuaFunctionCall in src/extractors/lua.ts:131 (2 transitive callers)
• handlePhpFuncCall in src/extractors/php.ts:304 (2 transitive callers)
• handlePhpMemberCall in src/extractors/php.ts:315 (2 transitive callers)
• handlePyCall in src/extractors/python.ts:163 (2 transitive callers)
• handleRubyCall in src/extractors/ruby.ts:154 (2 transitive callers)
• handleSwiftCallExpression in src/extractors/swift.ts:249 (2 transitive callers)
• printRoles in src/presentation/queries-cli/overview.ts:224 (2 transitive callers)

[carlos-alm]

Addressed both Greptile findings:

1. Orphaned JSDoc on runInterproceduralStitch (dataflow.ts lines 471-478): Removed the first stacked JSDoc block that described the old buildInterproceduralStitch wrapper semantics ("runs after all per-file vertices + summaries"). The second JSDoc ("Core stitch logic — must be called inside an already-open transaction") is the correct description for the internal helper and was kept.

2. PRAGMA foreign_keys = OFF scope too broad (native-orchestrator.ts): Added ctx.nativeDb.exec('PRAGMA foreign_keys = ON') immediately after buildGraph() returns, narrowing the FK-off window to just the Rust-side purge+build call. Post-build JS writes (gap-repair, structure patch) now have FK protection restored. The comment was updated to reflect this explicit restore.

[carlos-alm]

@greptileai

[greptile-apps]

If buildGraph() throws (native addon OOM, missing file, internal Rust panic, etc.), the PRAGMA foreign_keys = ON restore never executes. ctx.nativeDb lives in the shared PipelineContext and — given the comment explicitly calls out "gap-repair, structure patch" as subsequent writers — any outer catch block or pipeline teardown that reuses the connection would do so with FK enforcement still off. A try/finally bracket eliminates the gap.

Suggested change

	try {
	ctx.nativeDb.exec('PRAGMA foreign_keys = OFF');
	} catch {
	// exec may not exist on very old addon versions — safe to ignore
	}
	const resultJson = ctx.nativeDb.buildGraph(
	ctx.rootDir,
	JSON.stringify(ctx.config),
	JSON.stringify(ctx.aliases),
	JSON.stringify(ctx.opts),
	);
	// Restore FK enforcement immediately after buildGraph() so any subsequent
	// writes to this connection (gap-repair, structure patch) retain FK protection.
	try {
	ctx.nativeDb.exec('PRAGMA foreign_keys = ON');
	} catch {
	// safe to ignore on very old addon versions
	}
	try {
	ctx.nativeDb.exec('PRAGMA foreign_keys = OFF');
	} catch {
	// exec may not exist on very old addon versions — safe to ignore
	}
	let resultJson: string;
	try {
	resultJson = ctx.nativeDb.buildGraph(
	ctx.rootDir,
	JSON.stringify(ctx.config),
	JSON.stringify(ctx.aliases),
	JSON.stringify(ctx.opts),
	);
	} finally {
	// Restore FK enforcement so any subsequent writes to this connection
	// (gap-repair, structure patch) retain FK protection — even if buildGraph()
	// throws.
	try {
	ctx.nativeDb.exec('PRAGMA foreign_keys = ON');
	} catch {
	// safe to ignore on very old addon versions
	}
	}

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[carlos-alm]

Fixed — wrapped buildGraph() in try/finally so FK enforcement is always restored, even if the native call throws (OOM, Rust panic, etc.). The PRAGMA foreign_keys = ON restore is now in the finally block, mirroring the pattern you suggested.

[carlos-alm]

@greptileai