fix: scope incremental structure/roles neighbor expansion to the changed file

[carlos-alm]

Summary

Fixes the "1-file rebuild" benchmark regression (~85-155% vs the 3.15.0 baseline on both engines) that's currently blocking the pre-publish and perf-canary CI gates on every open PR, including #1885, #1886, and #1889.

Root cause: two independent, correct correctness fixes (#1840, #1849) each add cost to the incremental role/structure-metrics recompute that only shows up when the changed file is itself a "hub" (imported/re-exported by many other files) — exactly what the benchmark's fixed probe file, src/domain/queries.ts, is.

• fix: recompute directory structure metrics for affected directories on incremental rebuild #1840 — directory-metrics neighbor discovery (refresh_affected_directory_metrics / refreshAffectedDirectoryMetrics) was scoped to the changed file's whole containing directory, not the file itself, pulling in unrelated sibling files' edges. For queries.ts (leaf dir src/domain, itself a hub), this found 251 neighbour files / 29 affected directories, including the repo-root src/ (a near-full-table scan). Scoping to the exact file cuts this to 46 / 13.
• fix: scope reexportedSymbols to actually-named re-export specifiers #1849 — added symbol-level reexports edges so re-exported symbols get fan-in credit and aren't misclassified dead. That correctly widened do_classify_incremental's neighbour-file set for barrel files like queries.ts (0 → 15 files), which then hit two pre-existing inefficiencies never exercised at this scale on a 1-file change: an OR-based self-join SQLite can't index, and a repo-wide recursive CTE (the same one the full-classify path legitimately needs) answering "is this barrel production-reachable" for a question only 1-3 barrels actually need answered.

Fix (both native Rust and TypeScript/WASM)

• Scope neighbor discovery to the exact touched file (indexed point lookup) instead of a directory-range scan.
• Rewrite the OR self-join as a UNION of two directional equi-joins (identical result set, ~2.4x faster).
• Replace the global reachability closure with a backward-scoped check: find the small set of barrels reexporting into the affected files, then answer reachability per-barrel by walking reexports edges backward (bounded by that barrel's own chain depth) instead of materializing the whole graph's forward closure.

Every correctness guarantee from #1840/#1849 is preserved — verified against their dedicated test suites (issue-1738-structure-metrics-incremental.test.ts, issue-1742-reexport-symbol-scope.test.ts) on both engines.

Before / after (local, src/domain/queries.ts probe, native engine)

Metric	3.15.0 baseline	main (regressed)	This PR
1-file rebuild (native)	117ms	216-258ms (+85% to +120%)	139-166ms (+19% to +42%)
1-file rebuild (wasm)	112-113ms	240-286ms (+114% to +155%)	181-221ms (+60% to +97%, see note)
structureMs (native)	4ms	51-58ms	27-37ms
rolesMs (native)	18ms	42-58ms	17-23ms

Note: WASM's structureMs/rolesMs are fixed to the same degree as native (e.g. structureMs 44.7ms → 24-32ms, rolesMs 46.9ms → 16-23ms). Its total still shows some residual elevation on my local (shared, noisy) dev machine, traced to detectMs (change detection, unrelated to this fix — filed separately as #1948) rather than anything this PR touches.

Test plan

• Full test suite: 212 files, 3532 tests passed, 30 skipped (pre-existing), 2 todo (pre-existing)
• npm run lint clean
• cargo build --release / cargo check --workspace clean (only a pre-existing, unrelated warning)
• cargo test --release: 447 passed
• Dedicated regression suites for fix: recompute directory structure metrics for affected directories on incremental rebuild #1840 and fix: scope reexportedSymbols to actually-named re-export specifiers #1849 passing on both engines (18 + 103 tests)
• Engine-parity suite (tests/engines/, build-parity.test.ts, dropped-language-gap.test.ts) passing
• Repeated local benchmark measurements (both engines, multiple runs) confirming the structureMs/rolesMs phase-level fix is consistent, not a lucky sample — see perf: 1-file incremental rebuild regresses +79-81% (117ms -> ~209ms), likely from PR #1840 structure-metrics recompute #1855 for full numbers and reconciliation with CI: '1-file rebuild' benchmark baseline (117ms) appears stale against current runner performance #1864

Fixes #1855

[github-actions]

Codegraph Impact Analysis

4 functions changed → 5 callers affected across 3 files

• refreshAffectedDirectoryMetrics in src/domain/graph/builder/stages/build-structure.ts:297 (3 transitive callers)
• findDirectReexportBarrels in src/features/structure.ts:931 (2 transitive callers)
• isBarrelProdReachable in src/features/structure.ts:956 (2 transitive callers)
• classifyNodeRolesIncremental in src/features/structure.ts:1014 (1 transitive callers)

[greptile-apps]

Greptile Summary

This PR fixes a performance regression (~85–155% on the "1-file rebuild" benchmark) introduced by two prior correctness fixes (#1840, #1849) that each added cost to the incremental structure/role classification path when the changed file is a widely-imported hub like src/domain/queries.ts. Every correctness guarantee from those fixes is preserved.

• Structure neighbor scoping (structure.rs, build-structure.ts): find_neighbor_files now executes a point lookup on the exact changed file path (file = ?) instead of a directory range scan, cutting neighbor discovery from 251 → 46 files for the benchmark probe.
• Roles UNION rewrite (roles.rs, structure.ts): the find_neighbour_files OR-based self-join (un-indexable) is replaced by a UNION of two directional equi-joins (~2.4x faster); and the global prod_reachable recursive CTE is replaced by a per-barrel backward-walk bounded to the small set of barrels (typically 1–3) that actually reexport into the affected files.

Confidence Score: 5/5

Safe to merge — all three optimizations preserve the result sets of the queries they replace, correctness is backed by the dedicated regression suites for #1840 and #1849, and the Rust/TypeScript implementations mirror each other faithfully.

The neighbor-scoping change is a straightforward predicate swap (range to equality) that can only reduce the working set. The UNION rewrite is semantically equivalent to the original OR-join — both directions are covered, UNION deduplicates the same way DISTINCT did, and parameter binding is correctly repeated four times. The backward-scoped is_barrel_prod_reachable correctly answers the same yes/no question as the global CTE, and UNION (not UNION ALL) in the recursive CTE guarantees termination on cycles. The cachedStmt additions address the prior review comment cleanly.

No files require special attention.

Important Files Changed

Filename	Overview
crates/codegraph-core/src/features/structure.rs	Scopes find_neighbor_files to a single file path (indexed equi-join) instead of a directory range scan; updates refresh_affected_directory_metrics to loop over touched files directly. SQL and Rust binding are correct.
crates/codegraph-core/src/graph/classifiers/roles.rs	Rewrites find_neighbour_files as a UNION of two equi-joins (4x parameter groups correctly bound), introduces find_direct_reexport_barrels and per-barrel is_barrel_prod_reachable with prepare_cached, and replaces the global prod_reachable CTE in do_classify_incremental. Logic and SQL parameter binding verified correct.
src/domain/graph/builder/stages/build-structure.ts	TypeScript mirror of the structure.rs change: neighborFiles statement now uses named @file parameter instead of @lo/@hi range; loop iterates over individual files. Correct.
src/features/structure.ts	TypeScript mirror of the roles.rs changes: adds findDirectReexportBarrels, isBarrelProdReachable with module-level WeakMap caches, and updates classifyNodeRolesIncremental with the UNION neighbour query and scoped barrel check.

_{Reviews (3): Last reviewed commit: "fix: cache the per-barrel reachability s..." | Re-trigger Greptile}

[carlos-alm]

@greptileai