The builder walked tests/**/*.cktest only and emitted a code-gen pair from
every fixture's first page regardless of validity — so it trained Gemma on
deliberately-malformed snippets as canonical CkLang and never read the real
tests/corpus/** apps at all. Rewrite around two strictly-separated sources:

  - tests/corpus/**       -> code-gen (whole valid module files, prompt names
                             app + role + siblings) and impl-from-header
                             (.ckh -> .ck pairs).
  - tests/**/*.cktest      -> diagnostics only (all pages, canonical format,
                             (none) kept as a negative example). Never code-gen;
                             corpus subtree skipped by the fixture walk.

Split keys whole corpus apps to one side so sibling-context prompts can't leak
across train/eval. eval.py counts impl-from-header in the code-gen metric so
the new kind isn't dropped from reporting. README rewritten to document the
two-source split and the original bug.

@@ -0,0 +45,4 @@
+
+# ~MAX_CHARS keeps a chunk comfortably under a 2048-token sequence (~4 chars/
+# token), so the CPT trainer rarely truncates a chunk mid-thought.
+MAX_CHARS = 6000

@@ -106,0 +125,4 @@
+    # --- Phase 1: CPT (continued pre-training on raw CkLang text) -----------
+    #
+    # Same LoRA adapter, raw next-token over docs + corpus source. NOTE: this
+    # adapts the attention/MLP projections only — it does NOT retrain