Stack columns on mobile (<=700px) and add (N foci) hint for multi-box findings

app.py

@@ -312,8 +312,8 @@ def build_demo() -> Any:
         )
         gr.HTML(ui.ribbon_html())
 
-        with gr.Row(equal_height=False):
-            with gr.Column(scale=4, min_width=380):
+        with gr.Row(equal_height=False, elem_classes="rr-row"):
+            with gr.Column(scale=4, min_width=380, elem_classes="rr-col"):
                 with gr.Group(elem_classes="rr-panel"):
                     gr.HTML(ui.card_head("1", "Chest X-ray"))
                     image_in = gr.Image(
@@ -361,7 +361,7 @@ def build_demo() -> Any:
                         label="Examples (clean draft / draft that misses a finding / draft that over-calls a finding / draft that misses an urgent finding)",
                     )
 
-            with gr.Column(scale=6):
+            with gr.Column(scale=6, elem_classes="rr-col"):
                 with gr.Group(elem_classes="rr-panel"):
                     gr.HTML(ui.card_head(None, "Image-grounded evidence", legend=True))
                     overlay_out = gr.Image(show_label=False, elem_id="overlay-out", height=520, interactive=False)

cxr_auditor/comparator.py

@@ -27,6 +27,15 @@ label is a canonical positive.
 - URGENT: any image-present positive finding whose label is on the urgent-review
   whitelist (``findings.URGENT_WHITELIST`` via ``findings.is_urgent``).
 
+Matching is label-level (set-based), not per-lesion: every rule above keys on the
+canonical label, so a label is judged present-or-absent once, no matter how many
+boxes the image model localized for it. A finding the model grounds at two
+spatially-distinct sites yields at most one MISSING, one UNSUPPORTED, and one
+URGENT entry for that label; a second, separate lesion carrying the same label
+that the draft happens to omit is not flagged on its own. The overlay still draws
+every distinct box (see ``render``), so the overlay can legitimately show more
+boxes than the audit lists or the findings table have rows for that label.
+
 The canonical ordering of ``CANONICAL_FINDINGS`` is used for every output list so
 the result is deterministic regardless of input ordering and free of duplicates.
 """
@@ -108,7 +117,11 @@ def _present_image_findings(image_findings: Iterable[ImageFinding]) -> dict[str,
     sentinel are excluded: only positive findings the image asserts as present
     participate in the audit. When the same positive label appears more than once
     with status ``PRESENT``, the first occurrence (in input order) wins, so its
-    box and evidence are the ones surfaced.
+    box and evidence are the ones surfaced. Indexing by label makes the audit
+    label-level: a second present box of the same label (a spatially-distinct
+    lesion) collapses into this one entry and is not compared against the draft on
+    its own, so a draft that omits only that second site raises no separate flag.
+    The overlay still draws every distinct box (see ``render``).
     """
     present: dict[str, ImageFinding] = {}
     for finding in image_findings:

cxr_auditor/render.py

@@ -32,6 +32,14 @@ the others. ``cluster_overlay_boxes`` merges spatially-overlapping boxes into on
 is drawn exactly once with the correct, order-independent color and a single
 combined label.
 
+The findings table and the audit panel are deduplicated by canonical label (the
+audit reasons about each label once; see ``comparator``), while the overlay keeps
+one box per distinct region. A label the model localizes at several separate
+sites therefore appears once in the table and panel but draws several boxes, so
+the overlay can legitimately show more boxes than the table or panel have rows.
+``findings_table_rows`` appends an "(N foci)" hint to such a label so a reader can
+reconcile the single table row with the several boxes on the image.
+
 Label collision avoidance
 -------------------------
 Distinct regions can still carry long labels at nearly the same height (for
@@ -541,6 +549,25 @@ _FLAG_UNSUPPORTED = "Unsupported claim"
 _FLAG_NONE = ""
 
 
+def _foci_counts(outcome: AuditOutcome) -> dict[str, int]:
+    """Count the distinct overlay regions drawn for each canonical label.
+
+    Mirrors the overlay pipeline (``cluster_overlay_boxes(categorize_image_findings(...))``)
+    so the count matches what the image actually shows: each cluster is one drawn
+    region, and a label is counted once per cluster it belongs to. A label the model
+    localizes at two separate sites scores 2; two near-coincident boxes that merge
+    into one cluster score 1; a present finding with no box scores 0 (it draws
+    nothing). ``findings_table_rows`` uses a count above one to append an "(N foci)"
+    hint, reconciling the single deduplicated row with the several boxes the overlay
+    draws.
+    """
+    counts: dict[str, int] = {}
+    for cluster in cluster_overlay_boxes(categorize_image_findings(outcome)):
+        for label in cluster.labels:
+            counts[label] = counts.get(label, 0) + 1
+    return counts
+
+
 def findings_table_rows(outcome: AuditOutcome) -> list[list[str]]:
     """Build the structured findings table rows.
 
@@ -549,7 +576,10 @@ def findings_table_rows(outcome: AuditOutcome) -> list[list[str]]:
     ``status`` is ``"Present"`` or ``"Absent"``, and ``audit_flag`` is a plain
     English phrase. Image findings are deduplicated by canonical label (the model
     can localize one label to several boxes; the table shows it once), so a
-    non-expert never sees the same finding listed repeatedly. The ``no_finding``
+    non-expert never sees the same finding listed repeatedly; when the overlay
+    draws several distinct boxes for one label, that label's single row carries an
+    "(N foci)" hint so the one row reconciles with the several boxes on the image.
+    The ``no_finding``
     sentinel is never listed beside positive findings - it is mutually exclusive
     with them - so a negative draft ("lungs are clear") adds no contradictory row;
     a genuinely clear study instead shows a single ``no_finding`` row.
@@ -567,6 +597,7 @@ def findings_table_rows(outcome: AuditOutcome) -> list[list[str]]:
 
     missing = set(outcome.result.audit.missing_findings)
     urgent = set(outcome.result.audit.urgent_review_flags)
+    foci_by_label = _foci_counts(outcome)
     seen_image_labels: set[str] = set()
     for finding in outcome.result.image_findings:
         if finding.status is not FindingStatus.PRESENT or not _is_positive(finding.finding):
@@ -580,7 +611,11 @@ def findings_table_rows(outcome: AuditOutcome) -> list[list[str]]:
             flag = _FLAG_MISSING
         else:
             flag = _FLAG_SUPPORTED
-        rows.append([display_name(finding.finding), _SOURCE_IMAGE, _status_word(finding.status), flag])
+        name = display_name(finding.finding)
+        foci = foci_by_label.get(finding.finding, 0)
+        if foci > 1:
+            name = f"{name} ({foci} foci)"
+        rows.append([name, _SOURCE_IMAGE, _status_word(finding.status), flag])
 
     unsupported = set(outcome.result.audit.unsupported_claims)
     for draft_finding in outcome.result.draft_findings:

reading_room_ui.py

@@ -312,6 +312,28 @@ table.rr-table td.dim {{ color: var(--rr-muted); }}
 .rr-pad {{ padding: 12px 16px; }}
 footer {{ opacity: 0.55; }}
 .rr-panel .block {{ background: transparent !important; }}
+
+/* ---- mobile (<=700px): stack the two columns and let the custom flex rows
+   wrap, so a phone never overflows horizontally. Scoped entirely to this media
+   query, so the desktop and tablet layout above is untouched. The custom header,
+   ribbon, and footer are non-wrapping flex rows whose nowrap children would
+   otherwise pin the container wider than the viewport and defeat column
+   stacking; here they wrap and their nowrap text relaxes. The evidence column
+   (rr-col) is forced to a full-width flex line with no min-width floor so the
+   left and right columns stack instead of sitting side by side. ---- */
+@media (max-width: 700px) {{
+  #rr-header {{ flex-wrap: wrap; }}
+  #rr-header h1 {{ white-space: normal; }}
+  #rr-header .chip {{ white-space: normal; }}
+  #rr-header .spacer {{ flex-basis: 100%; }}
+  #rr-ribbon {{ flex-wrap: wrap; }}
+  #rr-footer {{ flex-wrap: wrap; }}
+  #rr-footer .spacer {{ flex-basis: 100%; }}
+  .rr-head {{ flex-wrap: wrap; }}
+  .rr-head .title {{ white-space: normal; }}
+  .rr-row {{ flex-wrap: wrap !important; }}
+  .rr-col {{ flex: 1 1 100% !important; min-width: 0 !important; }}
+}}
 """