Merge pull request #17 from interscript/fix/plain-parallel-duplicate-semantics

fix: plain parallel blocks take the replace-tree's duplicate semantics

Author: ronaldtse

src/interscript/engine.py

@@ -17,7 +17,7 @@
 import re
 import unicodedata
 
-from .expr import expr_lookbehind, expr_max_length, expr_neg_lookbehind, expr_to_literal, expr_to_regex
+from .expr import expr_is_par_safe, expr_lookbehind, expr_max_length, expr_neg_lookbehind, expr_to_literal, expr_to_regex, is_plain_string
 
 
 class ExecutionError(ValueError):
@@ -27,9 +27,30 @@ class ExecutionError(ValueError):
 def _compile_parallel(subs: list[dict]) -> tuple[re.Pattern[str], dict[str, str]]:
     """Compile one parallel group: longest-pattern-first alternation
     with a named group per sub; lookaround guards for before:/after:.
-    Plain-string patterns additionally feed the casing maps."""
+
+    An all-plain group compiles to a replace tree in Ruby, where a
+    later duplicate from overwrites an earlier one; a guarded group
+    takes the megaregexp, where the earliest equal-key rule wins."""
+    guards = ("before", "after", "not_before", "not_after")
+    all_plain = not any(sub.get(g) for sub in subs for g in guards) and all(
+        sub.get("result") is not None
+        and expr_is_par_safe(sub["pattern"])
+        and expr_is_par_safe(sub["result"])
+        for sub in subs
+    )
+    drop: set[int] = set()
+    if all_plain:
+        last: dict[str, int] = {}
+        for i, sub in enumerate(subs):
+            if is_plain_string(sub["pattern"]):
+                last[expr_to_literal(sub["pattern"])] = i
+        for i, sub in enumerate(subs):
+            if is_plain_string(sub["pattern"]) and last[expr_to_literal(sub["pattern"])] != i:
+                drop.add(i)
     indexed = []
     for i, sub in enumerate(subs):
+        if i in drop:
+            continue
         pat = expr_to_regex(sub["pattern"])
         full = pat
         if sub.get("before"):

src/interscript/expr.py

@@ -300,6 +300,17 @@ def is_plain_string(expr: str) -> bool:
     return bool(re.fullmatch(r'"(?:[^"\\]|\\.)*"', expr.strip()))
 
 
+def expr_is_par_safe(expr: str) -> bool:
+    """Ruby's parallel tree path builds literal from/to strings; a
+    boundary-like token there raises ("Can't use boundary in a par
+    context") and the whole block falls back to the megaregexp."""
+    try:
+        toks = _scan(expr, "par")
+    except ValueError:
+        return False
+    return all(k not in ("boundary", "nwb", "anchor") for k, _ in toks)
+
+
 def expr_max_length(expr: str) -> int:
     """The Ruby runtime's parallel-selection key: Rule::Sub#max_length =
     from + before + after + not_before + not_after (+ priority), where a

tests/test_engine.py

@@ -303,3 +303,46 @@ def test_subst_boundary_treats_combining_marks_as_word_chars():
     and the schwa-killing rule fired — kaṁganā came out kṁganā."""
     e = _load("un-mar-Deva-Latn-2016")
     assert e.transliterate("कंगना") == "kaṁganā"
+
+
+def test_plain_parallel_duplicate_last_wins():
+    """Measured against Ruby: an all-plain parallel block compiles to a
+    replace tree where a later duplicate from overwrites an earlier one
+    (masm-mon lists sub "i" "й" ... sub "i" "и" to let the last win).
+    A guarded block takes the megaregexp path instead, where the
+    earliest equal-key rule that matches wins."""
+    tree = parse_imp(
+        'stage {\n  parallel {\n'
+        '    sub "i", "й"\n'
+        '    sub "ab", "X"\n'
+        '    sub "i", "и"\n'
+        '    sub "ab", "Y"\n'
+        '  }\n}\n'
+    )
+    e = Engine(tree)
+    assert e.transliterate("i ab") == "и Y"
+
+    tree2 = parse_imp(
+        'stage {\n  parallel {\n'
+        '    sub "i", "X", before: "a"\n'
+        '    sub "i", "Y"\n'
+        '  }\n}\n'
+    )
+    e2 = Engine(tree2)
+    assert e2.transliterate("ai") == "aX"
+    assert e2.transliterate("bi") == "bY"
+
+
+def test_par_unsafe_rule_forces_megaregexp_first_wins():
+    """Measured via Ruby on bgnpcgn-bal: one rule with a boundary in a
+    par-unsafe position makes the whole block fall back from the
+    replace tree (last duplicate wins) to the megaregexp (earliest
+    equal-key rule wins) — و maps to o there, not the later w."""
+    tree = parse_imp(
+        'stage {\n  parallel {\n'
+        '    sub "i", "X"\n'
+        '    sub "i", "Y"\n'
+        '    sub "z" + boundary, "B"\n'
+        '  }\n}\n'
+    )
+    assert Engine(tree).transliterate("i") == "X"