feat(adr-0194): labeled-container subject entity shape — 'Jar A has N' parses, wrong=0-proven (substrate) (#499)

GSM8K labels containers/regions with a trailing single-letter or short-numeric
label ('Jar A has 28 marbles', 'Section G has 10 cars', 'District 2 has 19
voters'); the initial-possession entity slot captured only 'Jar' and the label
broke the match. Adds a separate sibling pattern _INITIAL_HAS_LABELED_RE
(mirroring ADR-0136.S.4 localisation) that REQUIRES the label, so the global
_ENTITY is unchanged and bare subjects yield no duplicate candidate.

- Composes with ADR-0193 aggregate question: 'Jar A has 28 marbles. Jar B has
  12 marbles. How many marbles are there in total?' -> 40.0.
- 0 real-corpus metric flip (honest substrate): the one real multi-container
  aggregate additionally needs comparative + multiplicative + lowercase-ref.
- wrong=0 HOLDS full corpus (7,473 q); train_sample byte-identical 4/46/0;
  synthetic-registry capability-axis gate + G5 lane green; smoke 67 passed.
- Label bounded by the possession verb: multi-word nouns ('Jar Apple') do NOT
  match. wrong=0 held downstream by completeness + round-trip + disagreement.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

docs/decisions/ADR-0194-labeled-container-subject.md

@@ -0,0 +1,84 @@
+# ADR-0194 — Labeled-container subject entity shape
+
+**Status:** Proposed (implemented in this PR).
+**Extends:** [ADR-0136.S.4](./ADR-0136.S.4-novel-initial-form.md) (sibling-pattern
+localisation), [ADR-0123a](./ADR-0123a-inference-shape-synonym.md) (entity slot).
+**Composes with:** [ADR-0193](./ADR-0193-aggregate-existential-question-frame.md)
+(aggregate question frame).
+**Substrate: 0 real-corpus metric flip by design; the value is the entity-shape
+generalisation + proven composition with the aggregate question.**
+
+> **One line.** GSM8K labels containers/regions with a trailing single-letter
+> or short-numeric label ("Jar A has 28 marbles", "Section G has 10 cars",
+> "District 2 has 19 voters"). The initial-possession entity slot
+> (`_ENTITY = (?:[A-Z]\w+|[Tt]he\s+\w+)`) captures only "Jar" and then expects
+> the possession verb, so the label breaks the match and the statement parses
+> to nothing. This adds a separate sibling pattern that REQUIRES a label.
+
+---
+
+## 1. The gap
+
+Both reader paths reject the labeled subject:
+- the candidate parser's `_INITIAL_HAS_RE` (`extract_initial_candidates` → 0
+  candidates);
+- the recognizer's discrete_count anchor (proper-noun single-token subject →
+  0 anchors).
+
+"Jamie has 28 marbles" parses (1 candidate); "Jar A has 28 marbles" does not —
+purely because of the trailing label.
+
+## 2. Decision
+
+Add `_INITIAL_HAS_LABELED_RE` in `generate/math_candidate_parser.py`, consumed
+by a dedicated `_init_has_labeled_candidates` helper wired into
+`extract_initial_candidates`:
+
+```
+^<Noun> <label> (has|have|had|started) <value> [adj] [unit] [of/in/for/with …]
+   label = a single uppercase letter OR 1-2 digits
+```
+
+- **Sibling-pattern localisation** (mirrors ADR-0136.S.4's
+  `_INITIAL_HAS_INDEF_RE`): the global `_ENTITY` is unchanged for every other
+  path (operations, comparisons, questions).
+- **The label is required**, so a bare subject ("Jamie has 28 marbles") never
+  reaches this pattern and yields no duplicate candidate.
+- **The label is bounded by the following possession verb**, so a multi-word
+  noun does NOT match: "Jar Apple has 5 marbles" → no candidate ("Apple" is not
+  a single-letter label), "Box Set has 12 items" → no candidate.
+- Same value/unit tail and money normalisation as `_INITIAL_HAS_RE`.
+
+### Why this is safe (the firewall is the precondition)
+
+The label widening only makes a statement *parse* into an initial possession.
+wrong=0 is held downstream by the completeness guard (ADR-0191) + the
+round-trip filter + branch disagreement — a mis-parse leaves source quantities
+uncovered and refuses. Full-corpus verification: **wrong=0 HOLDS** (7,473 q);
+`train_sample` byte-identical **4/46/0**; the synthetic-registry capability-axis
+`wrong=0` gate and the G5 aggregate lane both stay green.
+
+## 3. Evidence
+
+- **Composes with ADR-0193:** "Jar A has 28 marbles. Jar B has 12 marbles. How
+  many marbles are there in total?" → **40.0**; three-container variant → 10.0.
+- **0 real-corpus flip** (honest framing): of the 3 real container-subject
+  problems under an aggregate question, the only multi-container aggregate
+  ("Jar A has 28 marbles. Jar B has 12 more marbles than jar A. Jar C has twice
+  as many as jar B. … altogether?") additionally requires **comparative-additive
+  + comparative-multiplicative + lowercase-reference** resolution. This is the
+  composition-wall lesson again: an entity-shape widening is necessary, not
+  sufficient.
+- **Tests:** `tests/test_labeled_container_subject.py` (labeled containers
+  parse; bare subjects yield no duplicate; multi-word nouns don't match a
+  label; composes with the aggregate question).
+
+## 4. Consequences & follow-ups
+
+- The labeled-container entity shape is banked; it becomes load-bearing the
+  moment **comparative reading** (the actual aggregate blocker — ADR-0131.G.2 /
+  `core-comparatives`) can compose `compare_additive`/`compare_multiplicative`
+  ops into an aggregate answer, and the **lowercase-reference** ("jar A" inside
+  a later clause) resolves to the labeled entity.
+- Compound enumeration and the recognizer-path labeled subject remain closed;
+  open them only if a serving need is proven under the firewall.

generate/math_candidate_parser.py

@@ -256,6 +256,26 @@ _INITIAL_HAS_INDEF_RE: Final[re.Pattern[str]] = re.compile(
     r"\s*\.?$"
 )
 
+# ADR-0194 — labeled-container subject: "Jar A has 28 marbles.",
+# "Section G has 10 cars.", "District 2 has 19 voters.". GSM8K labels
+# containers/regions with a trailing single-letter or short-numeric label
+# that the bare _ENTITY slot cannot absorb. Sibling to _INITIAL_HAS_RE that
+# REQUIRES the label, so it never duplicates the bare-subject candidate;
+# _ENTITY stays unchanged for every other path. The label is a single
+# uppercase letter OR 1-2 digits, bounded by the following possession verb
+# (so a multi-word noun like "Jar Apple" does NOT match — "Apple" is not a
+# single-letter label). Same value/unit tail as _INITIAL_HAS_RE. wrong=0
+# is held downstream (completeness + round-trip + disagreement).
+_INITIAL_HAS_LABELED_RE: Final[re.Pattern[str]] = re.compile(
+    r"^(?P<entity>[A-Z]\w+\s+(?:[A-Z]|\d{1,2}))\s+"
+    r"(?P<anchor>has|have|had|started)(?:\s+(?:up|with))?\s+"
+    rf"(?P<value>{_VALUE})"
+    r"(?:\s+(?:full|loose|empty|whole|broken|new|old|small|large|fresh|raw|flat))?"
+    r"(?:\s+(?P<unit>\w+))?"
+    r"(?:\s+(?:of|in|for|with)\s+.+)?"
+    r"\s*\.?$"
+)
+
 # ADR-0136.S.4 — Shape B: prepositional-prefix existential.
 # "In a building, there are a hundred ladies on the first-floor studying."
 # Sibling to _INITIAL_THERE_ARE_RE; prefix is "In a <place>" (not bare
@@ -557,6 +577,9 @@ def extract_initial_candidates(sentence: str) -> list[CandidateInitial]:
     # ADR-0136.S.4 — Shape A: "A <noun> has N <unit>" indefinite-article subject.
     out.extend(_init_has_indef_candidates(sentence))
 
+    # ADR-0194 — labeled-container subject: "Jar A has 28 marbles."
+    out.extend(_init_has_labeled_candidates(sentence))
+
     m2 = _INITIAL_THERE_ARE_RE.match(s)
     if m2 is not None:
         value_raw = m2.group("value")
@@ -1946,6 +1969,52 @@ def _init_has_indef_candidates(sentence: str) -> list[CandidateInitial]:
         return []
 
 
+def _init_has_labeled_candidates(sentence: str) -> list[CandidateInitial]:
+    """ADR-0194 — labeled-container subject: 'Jar A has 28 marbles.'
+
+    Sibling to the _INITIAL_HAS_RE block in extract_initial_candidates.
+    Entity is '<Noun> <label>' (label = single uppercase letter or 1-2
+    digits), preserved by _normalize_entity. REQUIRES a label, so a
+    bare subject ('Jamie has 28 marbles') never reaches here and yields
+    no duplicate candidate. Same value/unit resolution and money
+    normalization as the definite-subject path.
+    """
+    s = sentence.strip().rstrip(".")
+    m = _INITIAL_HAS_LABELED_RE.match(s)
+    if m is None:
+        return []
+    value_raw = m.group("value")
+    rv = _resolve_value(value_raw)
+    if rv is None:
+        return []
+    entity = _normalize_entity(m.group("entity"))
+    unit_raw = m.group("unit")
+    if rv.unit_override is not None:
+        resolved_unit: str = rv.unit_override
+    elif unit_raw is not None:
+        resolved_unit = _canonicalize_unit(unit_raw)
+    else:
+        return []
+    value, final_unit = _money_unit_normalization(rv.value, resolved_unit)
+    assert final_unit is not None
+    try:
+        return [
+            CandidateInitial(
+                initial=InitialPossession(
+                    entity=entity,
+                    quantity=Quantity(value=value, unit=final_unit),
+                ),
+                source_span=sentence,
+                matched_anchor=m.group("anchor"),
+                matched_value_token=value_raw,
+                matched_unit_token=unit_raw if unit_raw is not None else final_unit,
+                matched_entity_token=m.group("entity"),
+            )
+        ]
+    except Exception:
+        return []
+
+
 def _init_there_are_prefix_candidates(sentence: str) -> list[CandidateInitial]:
     """Shape B — prepositional-prefix existential: 'In a building, there are 100 ladies.'
 

tests/test_labeled_container_subject.py

@@ -0,0 +1,81 @@
+"""ADR-0194 — labeled-container subject entity shape.
+
+GSM8K routinely labels containers/regions with a trailing single-letter or
+short-numeric label: "Jar A has 28 marbles.", "Section G has 15 rows.",
+"District 2 has 19 voters.".  The initial-possession parser's entity slot
+(``_ENTITY = (?:[A-Z]\\w+|[Tt]he\\s+\\w+)``) captures only "Jar" and then
+expects the possession verb, so the label breaks the match and the statement
+produces no candidate.
+
+This adds a SEPARATE sibling pattern ``_INITIAL_HAS_LABELED_RE`` (mirroring
+ADR-0136.S.4's ``_INITIAL_HAS_INDEF_RE`` localisation) that REQUIRES a label,
+so it never duplicates the bare-subject main pattern.  The global ``_ENTITY``
+is unchanged.  wrong=0 is held downstream by the completeness guard
+(ADR-0191) + round-trip + branch disagreement — the label widening only makes
+a statement *parse*; a mis-parse leaves quantities uncovered and refuses.
+
+SUBSTRATE: 0 real-corpus metric flip (the one real multi-container aggregate,
+"Jar A has 28 marbles. Jar B has 12 more than jar A. Jar C has twice as many
+as jar B. ...altogether?", additionally needs comparative + multiplicative
+reading).  Its value is the entity-shape generalisation + that it composes
+with the ADR-0193 aggregate question (proven below).
+"""
+from __future__ import annotations
+
+import pytest
+
+from generate.math_candidate_parser import extract_initial_candidates
+from generate.math_candidate_graph import parse_and_solve
+
+
+# --- Labeled-container subjects now parse as initial possessions -----------
+@pytest.mark.parametrize("sentence,entity,value,unit", [
+    ("Jar A has 28 marbles.", "Jar A", 28.0, "marbles"),
+    ("Box B has 15 marbles.", "Box B", 15.0, "marbles"),
+    ("Section G has 10 cars.", "Section G", 10.0, "cars"),
+    ("District 2 has 19 voters.", "District 2", 19.0, "voters"),
+    ("Tank 1 has 40 liters.", "Tank 1", 40.0, "liters"),
+])
+def test_labeled_container_parses(sentence, entity, value, unit) -> None:
+    cands = extract_initial_candidates(sentence)
+    assert len(cands) == 1, f"expected one candidate for {sentence!r}"
+    c = cands[0]
+    assert c.initial.entity == entity
+    assert c.initial.quantity.value == value
+    assert c.initial.quantity.unit == unit
+
+
+# --- No duplicate candidates for bare (unlabeled) subjects -----------------
+def test_bare_subject_single_candidate_unchanged() -> None:
+    """The labeled pattern requires a label, so 'Jamie has 28 marbles'
+    still yields exactly one candidate (from the main pattern only)."""
+    cands = extract_initial_candidates("Jamie has 28 marbles.")
+    assert len(cands) == 1
+    assert cands[0].initial.entity == "Jamie"
+
+
+# --- The label must not swallow a following content word -------------------
+@pytest.mark.parametrize("sentence", [
+    "Jar Apple has 5 marbles.",     # 'Apple' is not a single-letter label
+    "Box Set has 12 items.",         # 'Set' is not a label
+])
+def test_multiword_noun_not_a_label(sentence: str) -> None:
+    cands = extract_initial_candidates(sentence)
+    assert cands == [], f"{sentence!r} must not parse as a labeled container"
+
+
+# --- Composes with the ADR-0193 aggregate question -------------------------
+def test_composes_with_aggregate_question() -> None:
+    res = parse_and_solve(
+        "Jar A has 28 marbles. Jar B has 12 marbles. "
+        "How many marbles are there in total?"
+    )
+    assert res.answer == 40.0, res.refusal_reason
+
+
+def test_composes_three_containers() -> None:
+    res = parse_and_solve(
+        "Jar A has 5 marbles. Jar B has 3 marbles. Jar C has 2 marbles. "
+        "How many marbles are there altogether?"
+    )
+    assert res.answer == 10.0, res.refusal_reason