Update handoff to 0.741; record the header-classification tangle as the open problem

Names the real obstacle behind gasimova s and 10890106 s titles: three
interacting heuristics (col_start, is_title_font, the block-gap
threshold) each compensating for the others errors, so patching any one
moves the others. Four attempts are recorded with their costs so the
next session does not repeat them. Also flags Contact fields as a
smaller untouched item.

Author: jimnoneill

calibration/APPROACH_A_HANDOFF.md

@@ -8,8 +8,8 @@ runbook for approach A specifically.
 ## STATUS (2026-07-16): approach A done, LOGIC-GAP tail done, 20/21
 
 Board: **20/21 acceptable, 12/13 numbered end-to-end, 13/13 logic-OK**, corpus
-w=0.976 rGlobal=0.835 rField=0.737, suite green (260 passed). Reference
-snapshot: `baselines/try_final.json`. 13/13 logic-OK means that given ideal
+w=0.976 rGlobal=0.835 rField=0.741, suite green (260 passed). Reference
+snapshot: `baselines/try_wmedian.json`. 13/13 logic-OK means that given ideal
 reading order the corrector is now correct on every numbered poster in the
 corpus; the single remaining failure (8228476) is reading order, not logic.
 
@@ -18,7 +18,7 @@ real regression in this very session (gasimova's title fell 1.000 -> 0.706
 while its rField still ROSE, because another field gained more):
 
     ~/myenv/bin/python calibration/diagnostics/field_audit.py \
-        calibration/baselines/after_track_a.json calibration/baselines/try_final.json
+        calibration/baselines/after_track_a.json calibration/baselines/try_wmedian.json
     ~/myenv/bin/python calibration/diagnostics/annotation_audit.py
 
 `field_audit.py` diffs all 177 fields of all 21 posters between two snapshots
@@ -55,6 +55,11 @@ regressions):
   matcher.
 - a5e83de size-aware line clustering + baseline-nearest word mapping + curly
   apostrophe normalization (42 to PASS; also rField +0.101 on 5128504).
+- f43ed24 LINE_MAX_GAP (a line ends at a five-em gap, keeping a corner logo and
+  a poster-ID badge out of the title) + rejoin a title split across lines
+  (4446908 title 0.706 -> 1.000).
+- 50a5deb page median font measured over text, not blocks (10890106 Study
+  design 0.595 -> 0.962).
 
 **Poster 42's annotation was fixed in the corpus repo**
 (`fairdataihub/posters-science-extraction-api`, commit 21d238d, pushed to
@@ -113,18 +118,29 @@ matters, but the poster is RTL and blocked on approach D anyway.
   cannot rejoin the halves. Buys the byline (+0.154), Background (+0.396) and
   correct affiliations for all seven authors. Fixing it properly means making
   the flatten preserve genuine columns instead of dissolving the whole band.
-- **Tried and rejected for it** (both in git history, do not re-run blind):
+- **Tried and rejected for it** (all in git history, do not re-run blind):
   classifying large top-zone text as title fragments cost 4519718 0.372 on its
   banner and AISec2025 0.401; scaling the block-gap threshold to local line
   height instead of the page median cost 10890106 its title (0.757 -> 0.495)
-  and Acknowledgements (0.954 -> 0.549).
+  and Acknowledgements (0.954 -> 0.549); keeping banner furniture out of
+  `col_start` dropped the board to 19/21 and made 10890106 WORSE, not better.
+  That last one diagnosed the real problem (see below), so it was worth the
+  detour, but do not simply retry it.
 - **10890106 is extraction, not annotation** (its annotation is clean; audited).
-  Its weak fields are fragmentation: no block clears the header cutoff at all
-  (a narrow "Abstract nr" badge drags `col_start` to 90.6, cutoff 145.0, while
-  the title starts at 284.4), so its 72pt title stays two blocks, and Study
-  design / Conclusions scatter across blocks the metric scores one at a time.
-  The header cutoff being derived from the topmost narrow block is the common
-  thread with gasimova's title; that is the next thing worth fixing properly.
+  Its worst fields were fragmentation. Study design is fixed (0.595 -> 0.962)
+  by 50a5deb, which measured the page's median font over text rather than over
+  blocks. Its title (0.757) is still two blocks: a narrow "Abstract nr" badge
+  drags `col_start` to 90.6, cutoff 145.0, while the title starts at 284.4, so
+  no block on the poster clears the header cutoff and the title merge cannot
+  fire. Do NOT fix this by widening the cutoff (tried; see above). The honest
+  read is that header classification is a tangle of three interacting
+  heuristics - `col_start`, `is_title_font`, and the block-gap threshold - each
+  compensating for the others' errors. It wants a redesign against the corpus,
+  not another patch. `field_audit.py` will tell you immediately if you have
+  moved anything you did not mean to.
+- **Contact fields score low across several posters** (10890106 0.282,
+  gasimova 0.280): contact blocks are not being emitted as their own section.
+  Untouched, likely tractable, and worth a look before the header redesign.
 
 ## TL;DR