fix(review): sample last page, document flags-vs-routing, add flag-path tests

Address PR #855 review (all non-blocking):

- classifier: _sample_indices now always includes the first AND last page (the
  old evenly-spaced sample missed the tail, e.g. last sampled index 95 on a
  100-page doc -- a scanned tail could be missed).
- classifier + metrics: document that flags are diagnostic and fire
  independently of routing (image_heavy on ANY page vs the ocr route needing a
  page FRACTION), so flag{image_heavy} is expected to exceed classified{ocr}.
- classifier: clarify the text-quality whitespace comment (caps at 12%) and note
  the image double-count approximation (min() caps coverage).
- tests: add the scanned (no text layer) and bad_text_layer (junk text over an
  image) flag paths, and a test pinning first/last-page sampling.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

nextcloud_mcp_server/document_processors/classifier.py

@@ -82,7 +82,8 @@ def _text_quality(text: str) -> float:
     whitespace_ratio = sum(c.isspace() for c in text) / len(text)
     mean_token_len = sum(len(t) for t in tokens) / len(tokens)
     overlong_frac = sum(len(t) > 20 for t in tokens) / len(tokens)
-    # Clean prose: ~15-20% whitespace, mean token ~4-6 chars, ~no overlong tokens.
+    # Caps at 1.0 from 12% whitespace (conservative; clean prose runs 15-20%),
+    # mean token ~4-6 chars, ~no overlong tokens.
     ws_score = min(whitespace_ratio / 0.12, 1.0)
     len_score = (
         1.0 if mean_token_len <= 10 else max(0.0, 1.0 - (mean_token_len - 10) / 15)
@@ -94,9 +95,13 @@ def _text_quality(text: str) -> float:
 def _sample_indices(page_count: int) -> list[int]:
     if page_count <= MAX_SAMPLED_PAGES:
         return list(range(page_count))
-    # Evenly spaced sample across the document.
-    step = page_count / MAX_SAMPLED_PAGES
-    return sorted({int(i * step) for i in range(MAX_SAMPLED_PAGES)})
+    # Evenly spaced sample that always includes the first AND last page, so a
+    # scanned tail on an otherwise-digital doc isn't missed. Rounding collisions
+    # just yield a slightly smaller (still bounded) sample.
+    last = page_count - 1
+    return sorted(
+        {round(i * last / (MAX_SAMPLED_PAGES - 1)) for i in range(MAX_SAMPLED_PAGES)}
+    )
 
 
 def classify_pdf(content: bytes) -> DocClassification:
@@ -122,6 +127,9 @@ def classify_pdf(content: bytes) -> DocClassification:
             for img in page.get_images(full=True):
                 for rect in page.get_image_rects(img[0]):
                     img_area += abs(rect.width * rect.height)
+            # Approximate: an image placed multiple times (tiled backgrounds) is
+            # double-counted, so img_area can exceed page_area -- the min() caps
+            # coverage at 1.0, which is all the scanned/digital split needs.
             coverage = min(img_area / page_area, 1.0)
             # A page that is mostly a raster image is a scan/photo: its content
             # (handwriting, stamps, figure text) is not fully in any text layer,
@@ -143,6 +151,11 @@ def classify_pdf(content: bytes) -> DocClassification:
     )
     ocr_frac = (sum(p.needs_ocr for p in pages) / sampled) if sampled else 0.0
 
+    # Flags are diagnostic signals, intentionally independent of the routing
+    # verdict: image_heavy fires if ANY page is image-heavy, while the OCR route
+    # needs a FRACTION of pages (OCR_PAGE_FRACTION). So a mostly-digital doc with
+    # one full-page photo is flagged image_heavy yet still routes "fast" -- the
+    # flag_total{image_heavy} count is expected to exceed classified{ocr}.
     flags: set[str] = set()
     if any(p.image_coverage >= IMAGE_COVERAGE_SCANNED for p in pages):
         flags.add("image_heavy")

nextcloud_mcp_server/observability/metrics.py

@@ -285,6 +285,9 @@ document_classified_total = Counter(
 )
 
 document_classifier_flag_total = Counter(
+    # Diagnostic flags, independent of the routing verdict: image_heavy fires if
+    # ANY page is image-heavy whereas the ocr route needs a fraction of pages,
+    # so flag{image_heavy} is expected to exceed classified{recommended_tier=ocr}.
     "astrolabe_document_classifier_flag_total",
     "Tier-0 classifier flags raised on documents",
     ["flag"],  # image_heavy | scanned | bad_text_layer

tests/unit/test_doc_classifier.py

@@ -84,3 +84,45 @@ def test_large_doc_is_sampled():
     c = clf.classify_pdf(_digital_pdf(pages=120))
     assert c.page_count == 120
     assert c.sampled_pages <= clf.MAX_SAMPLED_PAGES
+
+
+def test_sample_indices_includes_first_and_last_page():
+    idx = clf._sample_indices(100)
+    assert idx[0] == 0
+    assert idx[-1] == 99  # last page must be sampled (scanned-tail case)
+    assert len(idx) <= clf.MAX_SAMPLED_PAGES
+
+
+# --- flag paths --------------------------------------------------------------
+
+
+def _image_with_mashed_text_pdf(pages: int = 2) -> bytes:
+    # Full-page image with a junk (mashed/space-less) text layer over it -- a
+    # scan whose OCR'd text layer is unusable.
+    doc = pymupdf.open()
+    pix = pymupdf.Pixmap(pymupdf.csRGB, pymupdf.IRect(0, 0, 600, 850))
+    pix.clear_with(255)
+    img = pix.tobytes("png")
+    mashed = "01322234567mobileoutstandingresilienceacademicachievement " * 3
+    for _ in range(pages):
+        page = doc.new_page(width=595, height=842)
+        page.insert_image(page.rect, stream=img)
+        page.insert_text((50, 60), mashed)
+    data: bytes = doc.tobytes()
+    doc.close()
+    return data
+
+
+def test_scanned_flag_when_no_text_layer():
+    c = clf.classify_pdf(_full_page_image_pdf())
+    assert c.total_chars == 0
+    assert "scanned" in c.flags
+    assert c.recommended_tier == "ocr"
+
+
+def test_bad_text_layer_flag_on_image_with_junk_text():
+    c = clf.classify_pdf(_image_with_mashed_text_pdf())
+    assert c.total_chars > 0
+    assert c.mean_text_quality < clf.MIN_TEXT_QUALITY
+    assert "bad_text_layer" in c.flags
+    assert c.recommended_tier == "ocr"