engine: count pruned directories in the receipt's skip accounting

walkRepo returned filepath.SkipDir before reaching skips.count++, so an
ignored directory's files landed in no bucket: not files_seen, not
files_skipped. The metric counted only ignored FILES the walker reached --
9 on a tree with ~67k ignored files, naming none of them. A bad ignore glob,
the thing this metric exists to catch, is almost always a directory glob.

Count a pruned directory once, in a new dirs_skipped field, rather than
walking node_modules/ purely to size it. Each skipped_sample entry now names
the glob that matched it, so "why is this file missing?" is a lookup. To get
the pattern, matchAnyGlob is split into matchGlob, which returns it;
matchAnyGlob wraps it, so both glob lists still share one matcher.

Exclude the configured output dir: .enola/** is itself a directory glob, so
counting it would make dirs_skipped differ between a repo's first snapshot
and every later one, and land as a phantom delta in diff_snapshot.

No cacheVersion bump -- facts.jsonl is byte-identical across the fix, so no
cachecov entry and no golden regeneration. Adds the tests that were missing:
nothing asserted files_skipped and no golden fixture contains an ignored
directory, which is how this survived 96 bumps.

Author: dejo1307

ARCHITECTURE.md

@@ -574,7 +574,7 @@ After `generate_snapshot`, these are written to the output directory (default `.
 | `facts.jsonl` | Every extracted fact, one JSON object per line |
 | `insights.json` | Architectural insights with confidence scores |
 | `snapshot.meta.json` | Metadata including per-file content hashes for incremental updates, plus the full receipt fields |
-| `receipt.json` | The **snapshot receipt** — a compact manifest of what the graph was generated over (enola version, git ref + dirty status, a content-fingerprint snapshot ID, the extractor/explainer sets, ignore-glob hash, output-artifact hashes) and extraction-quality metrics (files seen/parsed/skipped, parse errors, coverage gaps). Read it via the `snapshot_receipt` tool. |
+| `receipt.json` | The **snapshot receipt** — a compact manifest of what the graph was generated over (enola version, git ref + dirty status, a content-fingerprint snapshot ID, the extractor/explainer sets, ignore-glob hash, output-artifact hashes) and extraction-quality metrics (files seen/parsed/skipped, directory trees pruned, parse errors, coverage gaps). Read it via the `snapshot_receipt` tool. |
 | `previous/` | The immediately-preceding snapshot, auto-rotated on each write — the `baseline='previous'` source for `diff_snapshot` |
 | `baseline/` | A snapshot pinned by `set_baseline`, preserved across re-snapshots — the default `diff_snapshot` baseline |
 
@@ -583,7 +583,9 @@ After `generate_snapshot`, these are written to the output directory (default `.
 `receipt.json` (and the same fields inside `snapshot.meta.json`) exists to answer *"what was this graph deterministic over, and how complete is it?"* — the trust question before an agent relies on an `impact_analysis` or a `diff_snapshot`. It serves two consumers:
 
 - **Provenance / audit.** enola version, git ref + dirty-tree status, the extractor/explainer sets actually used, a **config hash** (over the effective extractors/explainers/renderers/globs/output settings) and its narrower `ignore_glob_hash`, per-artifact output hashes, and a **snapshot ID** that is a *content fingerprint* (SHA-256 over the byte-stable fact serialization plus the version and config hash), not a random UUID — so re-running on identical inputs yields the same ID and it can key equivalence. Every hash value carries a `sha256:` prefix.
-- **The improvement loop.** Extraction-quality metrics — files seen vs. parsed vs. skipped, a parse-error count and sample, the count of heuristic (confidence < 1.0) insights, and the cross-repo coverage-gap / unresolved-edge rollup — give a machine-readable signal a consumer (a human, a `diff_snapshot`, or an agent improving enola itself) can poll to detect *thin extraction* (a missing detection, a bad ignore glob, a failing extractor) and turn it into targeted work. The same metrics appear as an **Extraction Quality** section in `llm_context.md`, so an agent reading the snapshot sees thin extraction without a tool call.
+- **The improvement loop.** Extraction-quality metrics — files seen vs. parsed vs. skipped, the number of directory trees pruned, a parse-error count and sample, the count of heuristic (confidence < 1.0) insights, and the cross-repo coverage-gap / unresolved-edge rollup — give a machine-readable signal a consumer (a human, a `diff_snapshot`, or an agent improving enola itself) can poll to detect *thin extraction* (a missing detection, a bad ignore glob, a failing extractor) and turn it into targeted work. The same metrics appear as an **Extraction Quality** section in `llm_context.md`, so an agent reading the snapshot sees thin extraction without a tool call.
+
+  The two skip counters mean different things, and a bad ignore glob is usually a *directory* glob. `files_skipped` counts ignored files the walker **visited** and dropped — those matched by a file glob like `**/*.test.ts`. An ignored directory is pruned whole (`filepath.SkipDir`), so its contents are never visited and appear in no count: it is tallied once, as one entry in `dirs_skipped`. Each `skipped_sample` entry names the glob that matched it, so *"why is this file missing from the graph?"* is a lookup rather than an investigation; directories appear there with a trailing slash.
 
 Because the receipt fields live in `snapshot.meta.json`, they ride into every pinned/`previous` baseline, and `diff_snapshot` reads them to add a **comparability guard**: it warns (above the delta) when the baseline and current snapshots were *not* generated over equivalent inputs — a different repo, enola version, extractor set, or ignore-glob set — since a diff across a mismatched extractor set would report every one of that language's facts as spurious churn. `compare_receipts` surfaces the same verdict plus the metric deltas directly.
 

internal/diff/receipt.go

@@ -54,6 +54,7 @@ func CompareReceipts(base, cur facts.SnapshotMeta) *ReceiptComparison {
 		delta("files_seen", base.FilesSeen, cur.FilesSeen),
 		delta("files_parsed", base.FilesParsed, cur.FilesParsed),
 		delta("files_skipped", base.FilesSkipped, cur.FilesSkipped),
+		delta("dirs_skipped", base.DirsSkipped, cur.DirsSkipped),
 		delta("parse_errors", base.ParseErrors, cur.ParseErrors),
 		delta("coverage_gaps", baseGaps, curGaps),
 		delta("unresolved_edges", baseUnresolved, curUnresolved),

internal/diff/receipt_test.go

@@ -114,4 +114,31 @@ func TestCompareReceipts(t *testing.T) {
 			t.Error("expected Identical=true for matching snapshot_id")
 		}
 	})
+
+	// A newly-ignored directory prunes a whole subtree. It reports as a dirs_skipped
+	// delta and NOT as a quality regression: pruning vendor/ is usually the operator
+	// doing the right thing, and files_seen falling is the signal that costs something.
+	t.Run("a pruned directory surfaces as a dirs_skipped delta", func(t *testing.T) {
+		cur := base
+		cur.SnapshotID = "bbb"
+		cur.DirsSkipped = 2
+
+		rc := CompareReceipts(base, cur)
+
+		var found *MetricDelta
+		for i := range rc.Deltas {
+			if rc.Deltas[i].Name == "dirs_skipped" {
+				found = &rc.Deltas[i]
+			}
+		}
+		if found == nil {
+			t.Fatalf("no dirs_skipped delta; got %v", rc.Deltas)
+		}
+		if found.Before != 0 || found.After != 2 || found.Delta != 2 {
+			t.Errorf("dirs_skipped delta = %+v, want before=0 after=2 delta=2", *found)
+		}
+		if len(rc.QualityRegressions) != 0 {
+			t.Errorf("pruning a directory is not a regression, got: %v", rc.QualityRegressions)
+		}
+	})
 }

internal/engine/engine.go

@@ -317,6 +317,7 @@ func (e *Engine) GenerateSnapshot(ctx context.Context, repoPath string, appendMo
 			FilesSeen:         len(files),
 			FilesParsed:       e.store.CountFilesWithFacts(files, parsedPrefix),
 			FilesSkipped:      skips.count,
+			DirsSkipped:       skips.dirCount,
 			SkippedSample:     skips.sample,
 			IgnoreGlobHash:    ignoreGlobHash,
 			ParseErrors:       len(parseErrs),
@@ -442,12 +443,29 @@ func (e *Engine) flagUnmatchedRoutes() {
 	}
 }
 
-// walkSkips is a lightweight tally of files dropped by the ignore globs, kept so
-// a snapshot receipt can report how much of the tree was excluded (and a sample
-// of what) without retaining every skipped path.
+// walkSkips is a lightweight tally of what the ignore globs dropped, kept so a
+// snapshot receipt can report how much of the tree was excluded (and a sample of
+// what) without retaining every skipped path.
+//
+// Files and directories are tallied separately because they cost differently to
+// know. An ignored directory is pruned whole — the walker never descends, so its
+// contents are counted nowhere. Counting them would mean walking node_modules/
+// purely to size it, a stat per file for a number no architecture graph wants.
+// One pruned directory is one architecturally meaningful fact; its 55,041 files
+// are not.
 type walkSkips struct {
-	count  int
-	sample []string
+	count    int      // ignored FILES the walker visited
+	dirCount int      // ignored DIRECTORIES pruned; their contents are never visited
+	sample   []string // capped sample of both, each annotated with the glob that matched
+}
+
+// record appends "<path> (glob: <pattern>)" until the sample is full. Directories
+// arrive with a trailing slash, which is what distinguishes a pruned subtree from
+// a single dropped file in the receipt.
+func (s *walkSkips) record(path, pattern string) {
+	if len(s.sample) < skippedSampleCap {
+		s.sample = append(s.sample, path+" (glob: "+pattern+")")
+	}
 }
 
 // skippedSampleCap bounds the number of skipped paths retained for the receipt.
@@ -456,8 +474,8 @@ const skippedSampleCap = 20
 // walkRepo collects all files in the repo, applying ignore patterns. It returns
 // the indexable source files, separately the test/spec files matched by
 // TestGlobs (excluded from normal indexing but collected for reference-only
-// extraction — see runTestRefExtractors), and a tally of ignored files for the
-// snapshot receipt.
+// extraction — see runTestRefExtractors), and a tally of what the ignore globs
+// dropped — ignored files, and pruned directories — for the snapshot receipt.
 func (e *Engine) walkRepo(repoPath string) (files, testFiles []string, skips walkSkips, err error) {
 	err = filepath.WalkDir(repoPath, func(path string, d fs.DirEntry, err error) error {
 		if err != nil {
@@ -470,8 +488,15 @@ func (e *Engine) walkRepo(repoPath string) (files, testFiles []string, skips wal
 		}
 
 		// Skip ignored paths
-		if e.isIgnored(relPath, d.IsDir()) {
+		if pattern, ok := e.ignoreMatch(relPath); ok {
 			if d.IsDir() {
+				// enola's own output directory is not part of the source tree.
+				// Counting it would make dirs_skipped differ between a repo's
+				// first-ever snapshot and every one after it, for no signal.
+				if relPath != e.cfg.Output.Dir {
+					skips.dirCount++
+					skips.record(filepath.ToSlash(relPath)+"/", pattern)
+				}
 				return filepath.SkipDir
 			}
 			// An ignored FILE that is a test/spec is not indexed as production
@@ -481,9 +506,7 @@ func (e *Engine) walkRepo(repoPath string) (files, testFiles []string, skips wal
 				testFiles = append(testFiles, relPath)
 			}
 			skips.count++
-			if len(skips.sample) < skippedSampleCap {
-				skips.sample = append(skips.sample, filepath.ToSlash(relPath))
-			}
+			skips.record(filepath.ToSlash(relPath), pattern)
 			return nil
 		}
 
@@ -503,7 +526,15 @@ func (e *Engine) matchesTestGlob(relPath string) bool {
 // matchAnyGlob reports whether a forward-slash path matches any of the patterns.
 // It is the single matcher behind both the ignore list and the test globs, so a
 // file the two lists disagree about cannot exist: an ignored file that stops being
-// a test necessarily stops being ignored. Supported forms:
+// a test necessarily stops being ignored.
+func matchAnyGlob(relPath string, patterns []string) bool {
+	_, ok := matchGlob(relPath, patterns)
+	return ok
+}
+
+// matchGlob returns the first pattern that matches relPath. The receipt records it
+// beside the skipped path, so "why is this file missing from the graph?" is a
+// lookup rather than an investigation. Supported forms:
 //
 //	vendor/**                 anchored directory prefix
 //	**/build/**               a directory named "build" at any depth
@@ -512,7 +543,7 @@ func (e *Engine) matchesTestGlob(relPath string) bool {
 //
 // The last form is the only one that constrains directory and filename together;
 // see matchDirScopedGlob for why the Ruby test globs need it.
-func matchAnyGlob(relPath string, patterns []string) bool {
+func matchGlob(relPath string, patterns []string) (string, bool) {
 	for _, pattern := range patterns {
 		// "<prefix>/**/<fileglob>". Handled first and exclusively: the branches
 		// below would match such a pattern only when exactly one directory sits
@@ -522,7 +553,7 @@ func matchAnyGlob(relPath string, patterns []string) bool {
 			prefix, fileGlob := pattern[:i], pattern[i+len("/**/"):]
 			if !strings.Contains(fileGlob, "/") {
 				if matchDirScopedGlob(relPath, prefix, fileGlob) {
-					return true
+					return pattern, true
 				}
 				continue
 			}
@@ -532,31 +563,31 @@ func matchAnyGlob(relPath string, patterns []string) bool {
 			if seg != "" && !strings.Contains(seg, "/") {
 				for _, part := range strings.Split(relPath, "/") {
 					if part == seg {
-						return true
+						return pattern, true
 					}
 				}
 			}
 		}
 		if strings.HasSuffix(pattern, "/**") {
 			dirPrefix := strings.TrimSuffix(pattern, "/**")
 			if relPath == dirPrefix || strings.HasPrefix(relPath, dirPrefix+"/") {
-				return true
+				return pattern, true
 			}
 		}
 		if m, err := filepath.Match(pattern, relPath); err == nil && m {
-			return true
+			return pattern, true
 		}
 		if strings.HasPrefix(pattern, "**/") {
 			sub := strings.TrimPrefix(pattern, "**/")
 			if m, err := filepath.Match(sub, filepath.Base(relPath)); err == nil && m {
-				return true
+				return pattern, true
 			}
 			if m, err := filepath.Match(sub, relPath); err == nil && m {
-				return true
+				return pattern, true
 			}
 		}
 	}
-	return false
+	return "", false
 }
 
 // matchDirScopedGlob reports whether relPath's basename matches fileGlob AND
@@ -602,6 +633,12 @@ func (e *Engine) isIgnored(relPath string, isDir bool) bool {
 	return matchAnyGlob(filepath.ToSlash(relPath), e.cfg.Ignore)
 }
 
+// ignoreMatch reports whether a path is ignored, and by which pattern. The walker
+// needs the pattern to record it in the receipt's skipped sample.
+func (e *Engine) ignoreMatch(relPath string) (string, bool) {
+	return matchGlob(filepath.ToSlash(relPath), e.cfg.Ignore)
+}
+
 // runExtractors detects applicable extractors and runs them. When cache is
 // non-nil, extractors implementing plugin.FileOwner have their facts reused
 // whenever the files they depend on are unchanged since the last snapshot.