Record assistant accuracy hardening

CHANGELOG.md

@@ -33,6 +33,13 @@
 - Fixed the narrow-screen geographic explorer cascade so the theme selector,
   map and results stack at their container width instead of retaining the
   generic section grid and clipping the map horizontally.
+- Recognized Dutch compound GIS terms such as `bodemdetails`,
+  `bevolkingsontwikkeling` and `perceeloppervlaktes` when selecting the
+  minimal grounded assistant context.
+- Kept supporting object counts in the canonical response evidence but omitted
+  them from the model prompt whenever more meaningful area, length, population
+  or score metrics exist. This prevents a model typo from competing with the
+  authoritative semantic measurement.
 
 ## Sprint 213-214 Cross-domain area profile (2026-07-16)
 

docs/CODEX_EXECUTION_LOG.md

@@ -30,6 +30,15 @@ Changed:
   `section > div` grid overruled the explorer's responsive `display: block`.
   The geographic layout is now excluded from that generic rule and has a
   regression assertion for its mobile stacking contract.
+- A final accuracy audit compared the DOV inventory and assistant wording
+  directly with PostGIS. All 1,159 soil features are non-empty, valid and
+  intersect Mol, and their transformed area is 11,448.35 ha. The model still
+  mistyped the supporting count as 1,150 in prose even though its canonical
+  context contained 1,159.
+- Added compound-word theme recognition so `bodemdetails` now selects only the
+  soil context instead of the complete project portfolio. Supporting object
+  counts are retained in `context_metrics` but withheld from the model prompt
+  when semantic measurements exist, eliminating that irrelevant failure mode.
 
 Validation evidence:
 - A read-only live production-chain probe with the 1,200-token setting
@@ -60,6 +69,13 @@ Validation evidence:
   official-count wording or unrelated theme. The 390 px explorer measured
   317 px for theme, map and results panels with no page overflow; desktop and
   widescreen map checks remained clean and browser logs contained no warnings.
+- The focused live `bodemdetails` regression at commit `59194e9` selected one
+  DOV dataset and six canonical metrics, retained the exact 1,159 object count
+  in API evidence and returned only the five source-correct hectare metrics
+  plus the 1949-1971/1:20,000 limitation in end-user prose.
+- The final complete readiness gate passed 730 backend tests, compilation,
+  documentation/fixture/contract audits, one Alembic head, frontend typecheck,
+  production build and shell-script checks.
 
 Next:
 - Roll out the five governed thematic rasters municipality by municipality to