feat(perception): record E39 R1 refinement

docs/13_LIDAR_WORKER_PRODUCT_AND_ROADMAP.md

@@ -718,6 +718,18 @@ Execution was strictly sequential through E35: E30 determined what E31 was
 allowed to change; E31 determined the E32 profile; E32 determined the E33
 runtime input; E32/E33 then bound the E34 layer; E32–E34 then bound E35.
 Exact nominal and degradation accounting are closed, so A8 is complete.
+
+R1 perception quality is now measured by two immutable Worker 006 results.
+E38 established the baseline at 82.2% presence, 81.5% geometry association,
+95.2% freshness and 14 high-severity failures. E39 result
+`e39-perception-refinement-2fd253940c9d9a2fd3b3237f3f0932f81a9f69741935d793771243d5779af464`
+used a fixed package-bound camera + LiDAR feature projection and improved
+presence and geometry association to 84.9% while reducing high-severity
+failures to 8. Freshness remained above target at 93.2%. Its development
+cross-validation exceeded 90% but did not predict the sealed result, so R1
+remains open. The next iteration is development-only grouped time/scene
+qualification with a source-coordinate-free representation; validation is not
+opened for item-level tuning.
 E36 is the first generalization gate. A separate product decision follows:
 either keep the result as operator/shadow evidence, or start L5 occupied-space
 integration. No LAB in this cycle can enable navigation, commands or safety

docs/16_ARCHITECTURE_AUDIT_EXECUTION_ROADMAP.md

@@ -246,9 +246,18 @@ The first R1 baseline is E38 immutable Worker 006 result
 Its sealed 146-item validation measures presence at 82.2%, geometry
 association at 81.5% and freshness at 95.2%, with 100% accounting, zero
 false-free claims and 14 high-severity failures. R1 therefore remains open;
-the next bounded iteration targets detector/background presence and
+the next bounded iteration targeted detector/background presence and
 geometry-only association without changing the E37 validation set.
 
+E39 completed that refinement in immutable Worker 006 result
+`e39-perception-refinement-2fd253940c9d9a2fd3b3237f3f0932f81a9f69741935d793771243d5779af464`.
+Five-fold development CV crossed 90% for all three dimensions without
+validation-label access. Sealed validation improved presence to 84.9%,
+geometry association to 84.9% and reduced high-severity failures from 14 to
+8; freshness remained above target at 93.2%. The aggregate gate remains
+failed. R1 now advances to a harder grouped time/scene development protocol
+and a source-coordinate-free representation before any new sealed evaluation.
+
 - [x] Reproduce all 4,489 immutable E29 frames with the exact frozen profile
   before applying E31/E30 changes.
 - [x] Apply only the admitted semantic self-mask, two exact geometry

docs/20_RAVNOVES00_REFERENCE_SOURCE_PRODUCT_PLAN.md

@@ -78,6 +78,22 @@ high-severity validation cases fail, so R1 is not accepted. The next R1
 iteration is bounded to detector/background presence, geometry-only
 association and the camera-only freshness tail without changing validation.
 
+The second R1 measurement is immutable result
+`e39-perception-refinement-2fd253940c9d9a2fd3b3237f3f0932f81a9f69741935d793771243d5779af464`.
+Its fixed robust three-neighbour refinement used exact camera-crop, LiDAR
+shape and projection features and was selected only through five-fold
+development cross-validation. Development reached 90.3% presence, 90.6%
+geometry association and 96.5% freshness without validation-label access.
+Sealed validation reached 84.9%, 84.9% and 93.2% respectively. Accounting
+remains 100%, false-free claims remain zero and high-severity failures fall
+from 14 to 8, but R1 is still not accepted.
+
+The next R1 iteration must address the development-to-validation gap before
+adding model complexity: grouped time/scene development folds replace
+item-hash folds, and a source-coordinate-free representation is compared under
+that harder protocol. No individual E39 validation label may be used for
+diagnosis, fitting or selection.
+
 ## Experiment rules
 
 1. RAVNOVES00 remains immutable and is never relabelled as a LAB result.