feat/rag-intake — stop storing a compacted block as one averaged vector

.env.example

@@ -163,23 +163,41 @@
 # ever asked, so they compress the whole scale: hits and misses both
 # move closer together.
 #
-# So the fixtures validate the mechanism, not the number. On the real
-# store the number is nearer 0.96 -- from ONE pair, which is not enough
-# to set it. Re-run --no-plant with at least three questions on each
-# side (the harness now refuses fewer) before writing a value here.
+# So the fixtures validate the mechanism, not the number. Measured
+# again on 2026-08-22, six questions against a copy of the real store,
+# before and after it was re-sliced one entry per exchange:
 #
-# Read this before switching it on, though. The five rows behind the
-# invented-causality answer of 2026-08-19 sat at 0.90 to 1.0015 --
-# INSIDE the range a real hit occupies. No cutoff that keeps real hits
-# would have prevented that answer. This setting removes the case where
-# nothing in the store is remotely relevant; it does nothing about
-# several middling entries being welded together, which is the failure
-# that actually happened and which is an intake problem.
+#     before   hits 0.8934 0.671 0.9386 | misses 0.9891 0.9356 1.098
+#              worst hit 0.9386 > best miss 0.9356  ->  NO GAP
+#     after    hits 0.4498 0.671 0.7695 | misses 1.0619 0.8337 1.0369
+#              worst hit 0.7695 < best miss 0.8337  ->  gap 0.0642
 #
-# Re-measure against a COPY of your real store before trusting it:
+# The intake change is what made a threshold choosable at all. The
+# short fact (0.671) did not move -- only the entries that had been
+# buried in a compacted block did, which is the control.
+#
+# 0.95, not the midpoint the harness suggests. Two of the three misses
+# are "nothing here is relevant" (1.0619, 1.0369) and that is what this
+# setting is for; the third sits at 0.8337 only because the store
+# contains a lexically similar sentence that does not answer the
+# question, and no number separates that cleanly. Too high lets a bad
+# answer through, which gets argued with; too low answers "je n'ai rien
+# en mémoire" while holding the answer, which gets believed.
+#
+# recall.dropped logs every cut with its id and distance, so a value
+# that bites in the wrong place shows up in the logs.
+#
+# Re-measure against a COPY of your own store before trusting it:
 #     recall_distance.py --db /tmp/real_copy.db --no-plant \
 #         --hit "..." --miss "..."
-# RECALL_MAX_DISTANCE=1.05
+#
+# The one ACTIVE line in this file, deliberately. Everything else here
+# is a commented default that also lives in config.py; this value does
+# not, because it was measured against one specific store rather than
+# derived from anything general. Unset (the config default) still means
+# no filtering at all, so a deployment that has not measured its own
+# store is not silently given someone else's number.
+RECALL_MAX_DISTANCE=0.95
 
 # --- Memory ---------------------------------------------------------------
 # MEMORY_ENABLED=true