Measure the page's median font over text, not over blocks

The median font of a page is meant to say how big ordinary body text is;
every header heuristic is stated as a multiple of it. It was taken over
blocks, one vote each, so a crowd of tiny figure labels and axis ticks
outvoted the body. 10890106 lands on 6.1pt against a 72pt title, which
makes "title font" (1.4x median = 8.5pt) true of every block on the
poster: the test says nothing at all there. Weighting each block by its
length measures the body, which is what the multiples were chosen
against.

This is the root cause behind three patches attempted and rejected first
(all in the history, none worth re-running blind): admitting large
top-zone text as title fragments, scaling the block-gap threshold to
local line height, and keeping banner furniture out of col_start. The
last one is the tell - it made 10890106 WORSE, because a tight col_start
was the only thing holding a meaningless title-font test in check.
Widening the cutoff let the byline and legend become headers too. Fixing
the measure removes the need to compensate for it.

10890106 Study design 0.595 to 0.962, and its rField finally moves
(+0.029) after several attempts that could not touch it. 4552067 +0.096,
gasimova +0.013. Against that AISec2025 gives up 0.038, on a GT section
labelled "Conclusion" that holds a GitHub repo name, its URL and a
figure caption; the repo name now heads its own block, away from the
URL. Corpus rField 0.737 to 0.741 (0.729 at session start), w 0.976,
rGlobal 0.835, 20/21, 12/13 numbered, 13/13 logic-OK, suite green (260
passed), every affiliation string in the corpus unchanged and verified.

Unchanged: gasimova's title (0.706), whose cause is _flatten_top_band
dissolving a real two-column banner, not this measure.

Author: jimnoneill

calibration/baselines/try_final.json calibration/baselines/try_wmedian.jsoncalibration/baselines/try_final.json renamed to calibration/baselines/try_wmedian.json

@@ -19,7 +19,7 @@
       "id": "10890106",
       "w": 0.948,
       "r_global": 0.684,
-      "r_field": 0.713,
+      "r_field": 0.742,
       "fields": [
         [
           "title",
@@ -43,7 +43,7 @@
         ],
         [
           "S:Study design",
-          0.595,
+          0.962,
           26
         ],
         [
@@ -53,7 +53,7 @@
         ],
         [
           "S:Results: characterization MUTZ-3 d",
-          0.625,
+          0.61,
           55
         ],
         [
@@ -545,7 +545,7 @@
       "id": "4552067",
       "w": 0.994,
       "r_global": 0.97,
-      "r_field": 0.777,
+      "r_field": 0.873,
       "fields": [
         [
           "title",
@@ -554,7 +554,7 @@
         ],
         [
           "authors+affiliations",
-          0.025,
+          1.0,
           1
         ],
         [
@@ -569,7 +569,7 @@
         ],
         [
           "S:Network of Experts",
-          0.968,
+          0.859,
           77
         ],
         [
@@ -1011,7 +1011,7 @@
       "id": "AISec2025-poster",
       "w": 0.987,
       "r_global": 0.908,
-      "r_field": 0.811,
+      "r_field": 0.773,
       "fields": [
         [
           "title",
@@ -1020,12 +1020,12 @@
         ],
         [
           "authors+affiliations",
-          0.636,
+          0.696,
           15
         ],
         [
           "S:Introduction",
-          0.707,
+          0.873,
           148
         ],
         [
@@ -1040,12 +1040,12 @@
         ],
         [
           "S:Conclusion",
-          0.857,
+          0.4,
           12
         ],
         [
           "S:Selected papers citing CyberLLMIns",
-          0.969,
+          0.936,
           128
         ]
       ],
@@ -1178,7 +1178,7 @@
       "id": "gasimova(oos)",
       "w": 0.966,
       "r_global": 0.9,
-      "r_field": 0.784,
+      "r_field": 0.797,
       "fields": [
         [
           "title",
@@ -1217,12 +1217,12 @@
         ],
         [
           "S:Conclusion",
-          0.686,
+          0.883,
           106
         ],
         [
           "S:Contact",
-          0.359,
+          0.28,
           30
         ]
       ],

poster2json/extract.py

@@ -927,8 +927,18 @@ def extract_text_with_pdfplumber(pdf_path: str) -> Optional[str]:
             if not all_blocks:
                 continue
 
-            all_fontsizes = [blk["fontsize"] for blk in all_blocks if blk["fontsize"] > 0]
-            page_median_fs = sorted(all_fontsizes)[len(all_fontsizes) // 2] if all_fontsizes else 0
+            # The median font of a page means the size of typical TEXT, not of
+            # a typical block. Taking it per-block let a crowd of tiny figure
+            # labels and axis ticks outvote the body: 10890106 lands on 6.1pt
+            # against a 72pt title, so "title font" (1.4x median = 8.5pt) is
+            # true of every block on the poster and the test says nothing.
+            # Weighting each block by its length measures the body instead.
+            weighted_fs = []
+            for blk in all_blocks:
+                if blk["fontsize"] > 0:
+                    weighted_fs.extend([blk["fontsize"]] * max(1, len(blk["text"])))
+            page_median_fs = (sorted(weighted_fs)[len(weighted_fs) // 2]
+                              if weighted_fs else 0)
 
             text_left = min(blk["hpos"] for blk in all_blocks)
             text_right = max(blk["hpos"] + blk["width"] for blk in all_blocks)