perf(native): skip backfill on clean incrementals + bench guard tuning

[carlos-alm]

Summary

• perf(native) — Gate backfillNativeDroppedFiles on result.isFullBuild || result.removedCount > 0. After fix(native): persist file_hashes for dropped/symbol-less files #1069 made backfill run on every successful orchestrator pass and fix(native): skip unsupported-extension files in detect_removed_files #1070 taught the orchestrator to skip unsupported-extension files, the JS side was still calling backfill unconditionally — wasting ~45ms per incremental on this repo (fs walk + 2 DB queries + WASM re-parse of all 48 unsupported-extension fixture files) repairing a gap that no longer exists. Local measurement: 1-file rebuild dropped from ~108ms to ~60ms.
• test(bench) — Bump regression threshold from 25% → 50% with empirical justification, prune 6 dead KNOWN_REGRESSIONS entries (3.9.0/3.9.1/3.9.2 — these only fire when their version is latest, which never happens with current committed history), and add a 3.10.0:No-op rebuild entry documenting the verified ~3-7ms regression and its likely contributors (fix(native): skip unsupported-extension files in detect_removed_files #1070 filter, fix(native): persist file_hashes for dropped/symbol-less files #1069 file_hashes growth, tree-sitter 0.24→0.25).

Closes #1075.

Test plan

• CI bench gate passes (no-op + 1-file rebuild within threshold)
• npm run lint && npm run build
• RUN_REGRESSION_GUARD=1 npm run test:regression-guard passes locally (modulo the pre-existing wasm Full build delta against historical 3.9.4 data, which clears once 3.10.0 data lands)
• Verify backfill still runs on full builds and on incrementals with removed files

[claude]

Claude finished @carlos-alm's task —— View job

---

Root Cause Analysis

Problem Being Solved: Performance regression where backfillNativeDroppedFiles was called unconditionally on every native orchestrator pass, wasting ~45ms per incremental build to repair a gap that no longer exists after PRs #1069/#1070 improved native file handling.
Why This Approach: The fix correctly gates backfill on result.isFullBuild || result.removedCount > 0, eliminating unnecessary work on clean incrementals while preserving gap-repair behavior for cases that actually need it.
Risk Assessment: Low risk. The logic is conservative (still runs on full builds and when files are actually removed) and has clear rollback strategy.

Backlog Compliance

• Zero-dep: ✓ — No new dependencies added
• Foundation-aligned: ✓ — Pure performance optimization, maintains existing behavior contract
• Problem-fit: 4/5 — Directly reduces build latency for AI agents doing frequent incremental rebuilds
• Breaking: No — Purely internal optimization with identical external behavior
• Tier: 1

Code Quality Assessment

Performance Logic: ✅ The gating logic result.isFullBuild || removedCount > 0 is sound:

• removedCount is properly defaulted to 0 with nullish coalescing
• Full builds still get backfill (maintains engine parity)
• Incrementals only get backfill when files were actually removed
• Legacy binaries without the fix(native): skip unsupported-extension files in detect_removed_files #1070 filter still get the repair behavior

Type Safety: ✅ removedCount?: number in NativeOrchestratorResult is handled correctly with ?? 0

Documentation: ✅ Excellent comment explaining the optimization, referencing specific PRs and performance measurements

Test Changes: The regression guard updates are well-justified:

1. Threshold increase (25%→50%): Backed by empirical analysis of CI variance on sub-30ms metrics
2. Pruning stale entries: Removing 3.9.0/3.9.1/3.9.2 entries that only fire when their version is latest
3. Adding 3.10.0 entry: Documents verified small regression with root cause analysis

Critical Concerns

None identified. This is a clean performance optimization with proper safeguards.

Final Recommendation

• Rating: ⭐⭐⭐⭐☆ (4/5)
• Action: APPROVE
• Reasoning: Well-engineered performance fix addressing real overhead (~45ms savings per incremental). Logic is conservative, well-documented, and maintains backward compatibility. Regression guard updates are empirically justified rather than hiding issues.

---

Job Run | Branch: perf/skip-backfill-on-clean-incrementals

[greptile-apps]

Greptile Summary

This PR gates backfillNativeDroppedFiles on result.isFullBuild || removedCount > 0 || changedCount > 0, recovering ~45ms on clean incremental builds by skipping the expensive fs walk + 2 DB queries + WASM re-parse when the orchestrator reports no file activity. The bench guard is also refined: the global threshold stays at 25% while a new NOISY_METRICS set grants 50% tolerance only to sub-30ms metrics, and six dead KNOWN_REGRESSIONS entries are pruned.

• pipeline.ts: Guard broadened (from removals-only to removals-or-changes) to preserve correctness on incrementals with file activity, with the acknowledged residual gap for brand-new unsupported-extension files tracked in follow-up: residual backfill gap when new unsupported-extension file added on quiet incremental #1091.
• regression-guard.test.ts: NOISY_METRICS / thresholdFor() cleanly split tolerance by metric volatility; label strings match the checkRegression call sites exactly; 3.10.0:No-op rebuild entry correctly handles CI runs that exceed even the 50% noisy-metric floor.

Confidence Score: 5/5

Safe to merge — the backfill guard change is additive and the bench-guard refactor is internally consistent.

Both changed files are well-reasoned and self-consistent. The changedCount field is already defined and used in the same function, making the new guard type-safe. The NOISY_METRICS label strings were verified to match every checkRegression call site, and the two-layer tolerance strategy correctly catches mid-range regressions while tolerating documented sub-30ms jitter.

No files require special attention.

Important Files Changed

Filename	Overview
src/domain/graph/builder/pipeline.ts	Guard broadened to `isFullBuild
tests/benchmarks/regression-guard.test.ts	Adds NOISY_METRICS / NOISY_METRIC_THRESHOLD / thresholdFor() to give per-metric tolerance; label strings match all checkRegression call sites; dead 3.9.x KNOWN_REGRESSIONS entries pruned; 3.10.0 entry added for CI jitter above the 50% noisy-metric floor.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[tryNativeOrchestrator completes] --> B{result.isFullBuild?}
    B -- yes --> E[backfillNativeDroppedFiles]
    B -- no --> C{removedCount > 0?}
    C -- yes --> E
    C -- no --> D{changedCount > 0?\nnew in this PR}
    D -- yes --> E
    D -- no --> F[Skip backfill\n~45ms saved on clean incremental]
    E --> G[closeDbPair + return timing]
    F --> G

Loading

_{Reviews (4): Last reviewed commit: "Merge branch 'main' into perf/skip-backf..." | Re-trigger Greptile}

[greptile-apps]

Backfill silently skipped when a new unsupported-extension file is added on an incremental

When a developer adds a new .hcl, .scala, or .swift file (any extension outside Rust's file_collector) on an incremental build that has no other removals, the Rust orchestrator reports removedCount=0 and changedCount=0 (it never sees the file). The guard therefore skips backfill, leaving the new file absent from the graph until the next full rebuild with no warning.

Pre-#1069, backfill ran only on full builds, so this gap existed then too — but #1069 intentionally closed it by running backfill on every incremental. This PR re-opens that gap for the "new unsupported file, zero Rust-side activity" case. The perf win is real, but the condition could be broadened to result.isFullBuild || removedCount > 0 || (result.changedCount ?? 0) > 0 to at least cover incrementals where any file activity was detected, reducing (though not fully closing) the window.

[carlos-alm]

Good catch — broadened the guard to fire when result.changedCount > 0 as well, so any orchestrator-detected file activity now triggers backfill. This narrows the gap meaningfully (most incrementals do have changedCount > 0).

The residual case you identified (brand-new unsupported-extension file added on an otherwise-quiet incremental, where the orchestrator sees neither removals nor changes) requires JS-side detection of files the Rust collector doesn't see — tracked as a follow-up in #1091. Fixed in 6ace213.

[greptile-apps]

The 25%→50% bump is well-justified for sub-30ms noisy metrics, but paired with MIN_ABSOLUTE_DELTA = 10 it creates a blind spot for metrics in the 30–100ms range. At 50ms baseline, a 24ms regression (+48%, delta 24ms) would now pass silently where it would previously be caught. Consider leaving the threshold at 25% but adding a separate per-metric override (e.g. a NOISY_METRICS set) so only the genuinely high-variance sub-30ms entries get the wider tolerance.

Suggested change

	const REGRESSION_THRESHOLD = 0.5;
	const REGRESSION_THRESHOLD = 0.25;

[carlos-alm]

Agreed — kept REGRESSION_THRESHOLD at 0.25 so the 30-100ms range stays guarded, and added a NOISY_METRICS set with a separate 0.50 tolerance for the genuinely high-variance sub-30ms metrics (No-op rebuild, 1-file rebuild). A 50ms->74ms (+48%) regression now correctly fails. Fixed in b42e107.

[carlos-alm]

@greptileai

[github-actions]

Codegraph Impact Analysis

1 functions changed → 5 callers affected across 5 files

• tryNativeOrchestrator in src/domain/graph/builder/pipeline.ts:604 (5 transitive callers)