fix: bundle rfb/spec/xvar.nlp by re-syncing vendored NLPPlus/data with engine

.github/workflows/test.yml

@@ -13,6 +13,13 @@ jobs:
     - uses: actions/checkout@v4
       with:
         submodules: recursive
+    - name: Verify vendored data is in sync with engine
+      # NLPPlus/data is a committed copy of the engine's data/ tree that the
+      # wheel ships and __init__.py copies into the working folder at runtime.
+      # If it drifts, compiler passes added to the engine (e.g. rfb/spec/xvar.nlp
+      # for _xVAR) are missing and analyzers fail to build. Fail fast on drift;
+      # re-sync with: rsync -a --delete nlp-engine/data/ NLPPlus/data/
+      run: diff -rq nlp-engine/data NLPPlus/data
     - uses: actions/setup-python@v4
       with:
         python-version: '3.10'

.github/workflows/update-nlp-engine.yml

@@ -58,6 +58,14 @@ jobs:
           before=$(git rev-parse HEAD:nlp-engine)
           git submodule update --remote --recursive nlp-engine
           after=$(git -C nlp-engine rev-parse HEAD)
+          # Keep the vendored rule-compiler spec (NLPPlus/data) in lockstep with
+          # the engine's data/ tree. The wheel ships this copy and __init__.py
+          # copies it into the working folder at runtime; if it drifts, newly
+          # added compiler passes (e.g. rfb/spec/xvar.nlp, which teaches the
+          # compiler _xVAR("attr")) are missing and analyzers fail to build even
+          # though the engine source is correct. Mirror with --delete so removed
+          # files are dropped too.
+          rsync -a --delete nlp-engine/data/ NLPPlus/data/
           echo "latest=$latest"       >> "$GITHUB_OUTPUT"
           echo "published=$published" >> "$GITHUB_OUTPUT"
           echo "before=$before"       >> "$GITHUB_OUTPUT"
@@ -87,7 +95,7 @@ jobs:
         if: steps.bump.outputs.changed == 'true'
         run: |
           set -euo pipefail
-          git add nlp-engine
+          git add nlp-engine NLPPlus/data
           # The submodule may already be at `after` on main (advanced by another
           # path), leaving nothing to commit -- but we still need the tag to fire
           # publish.yml. Commit only when there's an actual change.

NLPPlus/data/rfb/spec/analyzer.seq

@@ -3,7 +3,8 @@ tokenize		#
 line		# 
 pat	retok	# 
 pat	bigtok	# 
-pat	x_white	# 
+pat	x_white	#
+pat	xvar	# merge _xVAR("attr") into one nonliteral token
 pat	nlppp	# comment here
 pat	un_mark	# comment here
 pat	list	# 

NLPPlus/data/rfb/spec/bigtok.nlp

@@ -96,7 +96,7 @@ _soRULES [base layer=(_startMark)] <- \@ RULES [t] @@
 	rfanonlit(2)
 	single()
 @RULES
-_NONLIT [base] <- \_ _xWILD [s one match=(_xALPHA _xEMOJI)] @@
+_NONLIT [base] <- \_ _xWILD [s plus match=(_xALPHA _xEMOJI _xNUM)] @@	# 06/11/26 DD. plus + _xNUM: allow digits in names (_B2, _34, _C25g).
 
 @RULES
 _ARROW [base] <- \< \- @@

NLPPlus/data/rfb/spec/xvar.nlp

@@ -0,0 +1,21 @@
+###############################################
+# FILE:     XVAR.PAT
+# SUBJ:     Merge _xVAR("attr") / _xVAR(attr) into a single nonliteral token.
+# NOTE:     Runs after bigtok forms _NONLIT/_STR and whitespace is zapped,
+#           and before list collects parenthesized match lists, so the inner
+#           parens of _xVAR(attr) never reach the list grammar.  The merged
+#           token _xVAR("attr") then flows through as one restricted-wildcard
+#           match-list entry; Pat::modeMatch1 tests the node attribute at
+#           runtime.  The attribute may be quoted (_STR) or an unquoted bareword
+#           (_LIT/_NUM); Pat::xvarMatch strips the optional quotes either way.
+###############################################
+
+@POST
+	rfaxvar(1,3)
+	single()
+@RULES
+# Quoted attr:    _xVAR("attr")
+_NONLIT [base] <- _NONLIT \( _STR \) @@
+# Unquoted attr:  _xVAR(attr)  /  _xVAR(123)
+_NONLIT [base] <- _NONLIT \( _LIT \) @@
+_NONLIT [base] <- _NONLIT \( _NUM \) @@