perf(output): Defer line-count and markdown-delimiter scans with lazy render-context getters

createRenderContext eagerly ran calculateFileLineCounts and
calculateMarkdownDelimiter — two full scans over every packed file's
content (~5.5MB) — on every run, even though fileLineCounts is only
consumed by skill generation and markdownCodeBlockDelimiter only by the
markdown/skill templates. Memoized getters defer both scans, so the
default XML (and plain/JSON) path never pays for them. Output is
byte-identical (verified with cmp for xml and markdown styles); when a
template or packSkill touches the property, the same computation runs
once and is cached.

Why this clears the bar now when earlier rounds rejected it: this
round's tail profiling showed that on warm runs (token cache hit) the
metrics workers resolve instantly and git token tasks complete ~28ms
before produceOutput, so generateOutput is the sole main-thread
bottleneck at the tail — the scans are wall-visible, not hidden behind
the parallel metrics branch as previously assumed.

Benchmark (packing this repo, warm, quiet 4-core Linux, interleaved
ABBA pairs, node bin/repomix.cjs --quiet):

- batch 1 (20 pairs): median delta -21.7ms, mean -20.1ms, t=-3.51, 15/20 improved
- batch 2 (30 pairs): median delta -25.6ms, mean -18.0ms, t=-2.49, 19/30 improved
- batch 3 (40 pairs): median delta -40.3ms, mean -40.9ms, t=-6.71, 37/40 improved
- pooled (90 pairs): mean -28.6ms on a ~1045ms baseline = -2.7%

Isolated scan cost is ~11-16ms (warm); the larger e2e delta is
consistent with reduced allocation/GC pressure from dropping the
per-line match(/\n/g) array allocations at peak heap.

intent(perf-tuning): automated round targeting >=2% end-to-end CLI improvement with behavior preserved
decision(render-context): memoized getters over style-conditional skips — RenderContext shape and all call sites unchanged, and any consumer that does read the properties still gets identical values
rejected(file-search): gitignore:false + prebuilt globby predicate measured -5.2% e2e but unshippable — depends on globby's unexported ignore.js internals (exports-map patch won't exist for npm consumers) and loses globby's gitignore-pattern directory pruning, risking large traversal regressions on repos with big ignored dirs and no negation patterns
learned(metrics-tail): on warm runs the tail critical path is produceOutput -> write -> wrapper extraction; git diff/log token tasks finish ~28ms earlier, so render-side CPU cuts are wall-visible up to that slack

https://claude.ai/code/session_01RD8vNvv1qtYV8BgdxMU7js

Author: claude

src/core/output/outputGenerate.ts

@@ -76,6 +76,11 @@ const calculateFileLineCounts = (processedFiles: ProcessedFile[]): Record<string
 };
 
 export const createRenderContext = (outputGeneratorContext: OutputGeneratorContext): RenderContext => {
+  // fileLineCounts is only consumed by skill generation and markdownCodeBlockDelimiter only by the
+  // markdown/skill templates; both require a full scan over every file's content. Memoized getters
+  // defer that scan so the default XML (and plain/JSON) path never pays for it.
+  let fileLineCounts: Record<string, number> | undefined;
+  let markdownCodeBlockDelimiter: string | undefined;
   return {
     generationHeader: generateHeader(outputGeneratorContext.config, outputGeneratorContext.generationDate),
     summaryPurpose: generateSummaryPurpose(outputGeneratorContext.config),
@@ -89,12 +94,18 @@ export const createRenderContext = (outputGeneratorContext: OutputGeneratorConte
     instruction: outputGeneratorContext.instruction,
     treeString: outputGeneratorContext.treeString,
     processedFiles: outputGeneratorContext.processedFiles,
-    fileLineCounts: calculateFileLineCounts(outputGeneratorContext.processedFiles),
+    get fileLineCounts() {
+      fileLineCounts ??= calculateFileLineCounts(outputGeneratorContext.processedFiles);
+      return fileLineCounts;
+    },
     fileSummaryEnabled: outputGeneratorContext.config.output.fileSummary,
     directoryStructureEnabled: outputGeneratorContext.config.output.directoryStructure,
     filesEnabled: outputGeneratorContext.config.output.files,
     escapeFileContent: outputGeneratorContext.config.output.parsableStyle,
-    markdownCodeBlockDelimiter: calculateMarkdownDelimiter(outputGeneratorContext.processedFiles),
+    get markdownCodeBlockDelimiter() {
+      markdownCodeBlockDelimiter ??= calculateMarkdownDelimiter(outputGeneratorContext.processedFiles);
+      return markdownCodeBlockDelimiter;
+    },
     gitDiffEnabled: outputGeneratorContext.config.output.git?.includeDiffs,
     gitDiffWorkTree: outputGeneratorContext.gitDiffResult?.workTreeDiffContent,
     gitDiffStaged: outputGeneratorContext.gitDiffResult?.stagedDiffContent,