Merge nucleic/eager-glass-wren-mrb5 into dev

datagen-prompt.md

@@ -9,10 +9,15 @@ pasted into the generator's context (contamination).
 ---
 
 You are generating a labeled dataset of prompts that software developers type into a
-coding-agent app (like Claude Code) to start or continue a chat. Each example is ONE
-prompt a real developer might send, labeled with what the prompt is FOR. The data trains
-a small on-device classifier, so realism and diversity matter more than polish; label
-precision matters more than anything.
+coding-agent app (like Claude Code) to START a new chat. Each example is the OPENING
+message of a fresh chat — never a reply inside an ongoing conversation — labeled with
+what the prompt is FOR. This matters: the classifier runs exactly once, on the first
+message, to pick the chat's model (which then stays fixed for the chat's whole life), so
+the training distribution must be first-messages only. A first message may still
+reference prior work the way developers really do ("continuing from yesterday's auth
+refactor, …", "picking up the payment-flow bug again"), because people routinely start
+fresh chats mid-project. The data trains a small on-device classifier, so realism and
+diversity matter more than polish; label precision matters more than anything.
 
 ## Labels (choose the primary purpose; definitions are exhaustive)
 
@@ -64,9 +69,10 @@ Boundary rules (apply in this order when two labels tempt you):
   - `pasted-context` — the ask is buried in pasted material: stack traces, log tail,
     diff hunk, failing test output, a TODO list. Reproduce the pasted material
     realistically, 100–400 tokens (≈10%)
-  - `vague-eval` — terse/ambiguous prompts with no recoverable purpose ("continue",
-    "make it pop", "do the thing we discussed"). Label with your best guess anyway;
-    these are held out of training by the pipeline (≈5%)
+  - `vague-eval` — terse/ambiguous prompts with no recoverable purpose, typed as the
+    first message of a new chat ("continue", "make it pop", "do the thing we
+    discussed" — real behavior when someone reopens work in a fresh chat). Label with
+    your best guess anyway; these are held out of training by the pipeline (≈5%)
 
 ## Diversity requirements (enforced per batch)
 
@@ -97,6 +103,11 @@ Boundary rules (apply in this order when two labels tempt you):
   (except the `vague-eval` slice, where that's the point).
 - Assistant-directed meta ("classify this prompt", "as an AI") — these are prompts TO a
   coding agent, about code.
+- Mid-conversation replies: anything that only makes sense as turn 2+ of a thread —
+  reacting to an assistant's previous answer ("yes do option 2", "that didn't work, try
+  again", "same error as before", "looks good, ship it", "no, the OTHER function").
+  Referencing prior *work* in a fresh chat is fine (see the intro); referencing a prior
+  *turn of this conversation* is not, because the classifier never sees those.
 
 ## Process