feat(adr-0192): open discrete_count noun class — 8x statements parse, wrong=0-proven (substrate) (#497)

The discrete_count matcher gated the counted noun on a CLOSED ratified set
(observed_counted_nouns): 'Betty has 24 marbles' matched, 'Randy has 60 mango
trees' / 'Sam has 12 red apples' did not — purely because the noun was unseen.

Open the single-anchor possession/acquisition path to an open noun phrase
(adjective* + 1-3 word head, bounded by a stop-word lookahead so it never
swallows a trailing PP), keeping every other narrowness layer (proper-noun
subject, verb whitelist, single numeric token, no clause-split). Closed
observed nouns still match (capitalized compounds preserved); compound
enumeration stays closed.

Safe because ADR-0191 moved the wrong=0 guarantee downstream: an open-vocab
mis-parse hits the completeness guard + round-trip + branch-disagreement.
Proof: full real corpus 61->494 discrete_count anchors (8x), wrong=0 HOLDS,
zero confabulations.

Substrate PR — 0 metric delta by design (train_sample byte-identical 4/46/0;
the problems still need composition downstream). Value: the foundation every
discrete_count flip consumes, and empirical proof open-vocab is firewall-safe.

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

docs/decisions/ADR-0192-discrete-count-open-noun-class.md

@@ -0,0 +1,114 @@
+# ADR-0192 — Open the discrete_count counted-noun class (firewall-backed)
+
+**Status:** Proposed (implemented in this PR). Widens the
+[ADR-0163.D.2](./ADR-0163-recognizer-storage.md) discrete_count matcher.
+Builds directly on [ADR-0191](./ADR-0191-candidate-graph-completeness-guard.md)
+— the completeness firewall is the precondition that makes this safe.
+**Substrate PR: 0 metric delta by design; the value is 8× more statements
+parsing into solver state, wrong=0-proven on the full real corpus.**
+
+> **One line.** The discrete_count matcher gated the counted noun against a
+> CLOSED ratified set (`observed_counted_nouns`): "Betty has 24 marbles"
+> matched only because "marbles" was ratified, while "Randy has 60 mango
+> trees" / "Sam has 12 red apples" produced no anchor purely because the noun
+> was unseen. This opens the single-anchor possession/acquisition path to an
+> open noun phrase, keeping every other narrowness layer. Wrong=0 is held
+> downstream by the ADR-0191 completeness guard + round-trip + branch
+> disagreement — not by the curated noun list.
+
+---
+
+## 1. The gap (microscope finding, 2026-05-30)
+
+The full-corpus microscope (`scripts/gsm8k_microscope.py`) ranked the serving
+reader's refusals across all 7,473 real GSM8K train questions.
+**`discrete_count_statement` is the dominant wall: 3,850 first-wall refusals**
+("recognizer matched but produced no injection"). Dissecting *why* the matcher
+emits no anchor:
+
+| sub-shape | count | extractable? |
+|-----------|------:|--------------|
+| `subj verb N <multi-word / adj+noun>` ("Randy has 60 **mango trees**") | ~1,004 | **yes — matcher too narrow** |
+| count on a prepositional object ("sold clips **to 48** friends") | ~550 | no — correctly conservative |
+| attributive number ("a **120-page** book") | ~120 | no — verb not possession/acquisition |
+| number is a unit (rate/currency/time) | ~380 | no — different category |
+| relational / "other" | ~1,400 | no — needs composition |
+
+Pinned blocker: the matcher only extracts when the counted noun is in
+`spec.observed_counted_nouns` (a closed ratified set). `"Betty has 24
+marbles"` matched (ratified); `"Randy has 60 mango trees"` / `"Sam has 12 red
+apples"` / `"Randy has 60 trees on his farm"` all emitted **anchors=0** solely
+because the noun (or noun phrase) was unseen — not because of the trailing PP
+(the regex already allowed trailing content) and not because the shape was
+ambiguous.
+
+## 2. Decision
+
+Open the counted-noun slot of the **single-anchor** discrete_count extractor
+(`_extract_discrete_count_re_open` in `generate/recognizer_match.py`):
+
+- The noun slot matches either a ratified `observed_counted_nouns` entry
+  (closed branch — preserves casing canonicalization and capitalized
+  compounds like "Pokemon cards") **OR** an OPEN lowercase noun phrase:
+  1–3 consecutive lowercase word tokens, none a boundary/stop word
+  (prepositions, conjunctions, determiners, comparatives).
+- `(?-i:...)` makes the open branch lowercase-only so it never captures a
+  following proper noun; the stop-word lookahead bounds the phrase so it
+  never swallows a trailing prepositional phrase ("mango trees on his farm"
+  → "mango trees").
+- **Every other narrowness layer is unchanged**: proper-noun subject,
+  possession/acquisition verb whitelist, single numeric token, no
+  clause-split. The compound-enumeration path stays closed.
+
+### Why this is safe (the firewall is the precondition)
+
+The closed noun set existed to prevent open-vocabulary mis-parses from
+reaching the solver. ADR-0191 moved that guarantee downstream: an open-vocab
+mis-parse now hits the **completeness guard** (every source quantity must be
+consumed), the **round-trip filter** (every slot must ground in source), and
+**branch-disagreement** refusal. So wrong=0 is held by the firewall, not by
+the noun list. The dangerous shapes are still refused *before* the open noun
+even applies — `"is reading a 120-page book"` refuses because "is" is not a
+possession/acquisition verb; `"has many apples"` refuses on the count token;
+`"has 60 apples and 30 oranges"` refuses on the single-count / clause-split
+layers.
+
+## 3. Evidence
+
+- **Substrate gain: 61 → 494** discrete_count anchors extracted+injected over
+  the full real corpus (8×), all clean.
+- **wrong=0 holds** on the full 7,473-question corpus — 494 statements parse,
+  **zero confabulations**. This is the direct proof that open-vocabulary
+  recognition is safe under the ADR-0191 firewall.
+- **0 metric delta** (`train_sample` byte-identical **4/46/0**; full-corpus
+  correct unchanged at 4). The widening makes *statements* parse; the
+  *problems* still refuse downstream at the composition wall (multi-statement
+  chaining + question-target). This is expected: statement parsing is
+  necessary, not sufficient. Refusal families shift accordingly — problems
+  advance from the discrete_count first-wall to later walls.
+- **Tests:** new `tests/test_discrete_count_open_noun_class.py` (open-vocab
+  now extracts; noun phrase stops before prepositions; dangerous shapes still
+  refuse). The one closed-contract assertion
+  (`test_unobserved_counted_noun_refused`) is updated to the new open
+  contract. All other discrete_count narrowness tests unchanged and passing.
+
+## 4. Consequences
+
+- This is **substrate**, deliberately landed with no metric movement. Its
+  value is (a) the foundation every discrete_count composition will consume —
+  a statement cannot be composed before it parses — and (b) the empirical
+  proof that the firewall makes open-vocabulary recognition wrong=0-safe,
+  retiring the closed-set constraint for the simple possession/acquisition
+  shape.
+- The remaining discrete_count walls (prepositional-object counts,
+  attributive numbers, rate/currency) are correctly still refused — they are
+  *not* simple possession and must not be admitted by this path.
+- The next layer is composition (multi-statement same-unit aggregate +
+  question-target parsing) which now has parsing statements to consume.
+
+## 5. Follow-ups
+
+- Re-run `scripts/gsm8k_microscope.py --corpus <train.jsonl>` after the
+  composition layer lands to confirm wrong=0 holds *and* the metric moves.
+- Compound-enumeration ("N1 noun1 and N2 noun2") noun class remains closed;
+  open it only after the single-anchor open path is proven in serving.

generate/recognizer_match.py

@@ -895,6 +895,53 @@ def _extract_discrete_count_re_for(counted_nouns: list[str]) -> re.Pattern[str]:
     )
 
 
+# ADR-0192 — words that terminate (cannot be part of) an open counted-noun
+# phrase: prepositions, conjunctions, determiners, and comparative markers.
+# Bounding the phrase against these is what stops the open noun from
+# swallowing a trailing prepositional phrase ("mango trees on his farm" →
+# "mango trees", not "mango trees on his farm").
+_OPEN_NOUN_STOP: Final[str] = (
+    "on|in|at|to|for|with|of|from|by|per|into|onto|over|under|"
+    "and|or|but|than|as|that|which|who|whose|whom|while|when|because|"
+    "the|a|an|his|her|its|their|our|your|my|each|every|"
+    "more|fewer|less|most|fewest|other|another"
+)
+
+
+def _extract_discrete_count_re_open(counted_nouns: list[str]) -> re.Pattern[str]:
+    """ADR-0192 — open-vocabulary variant of the single-anchor extractor.
+
+    Strictly additive: the counted-noun slot matches either a ratified
+    ``observed_counted_nouns`` entry (closed branch — preserves casing
+    canonicalization and capitalized compounds like ``Pokemon cards``) OR
+    an OPEN lowercase noun phrase: 1–3 consecutive lowercase word tokens,
+    none a boundary/stop word.  The ``(?-i:...)`` makes the open branch
+    lowercase-only so it never captures a following proper noun, and the
+    stop-word lookahead bounds the phrase so it never swallows a trailing
+    prepositional phrase.  Every other narrowness layer (proper-noun
+    subject, verb whitelist, single numeric token, no clause-split) is
+    unchanged; wrong=0 is held downstream by the ADR-0191 completeness
+    guard + round-trip + branch-disagreement.
+    """
+    options = sorted({n for n in counted_nouns if n}, key=len, reverse=True)
+    closed_alt = "|".join(re.escape(n) for n in options)
+    open_tok = rf"(?-i:(?!(?:{_OPEN_NOUN_STOP})\b)[a-z]+)"
+    open_noun = rf"{open_tok}(?:\s+{open_tok}){{0,2}}"
+    noun_group = (
+        rf"(?P<noun>{closed_alt}|{open_noun})" if closed_alt
+        else rf"(?P<noun>{open_noun})"
+    )
+    return re.compile(
+        r"^\s*"
+        r"(?P<subject>(?-i:[A-Z][a-z]+))"
+        r"\s+(?P<verb>[A-Za-z]+)"
+        r"\s+(?P<count>\d+|[A-Za-z\-]+)"
+        r"\s+" + noun_group +
+        r"(?:\b.*)?$",
+        flags=re.IGNORECASE,
+    )
+
+
 _DIGIT_RUN_RE: Final[re.Pattern[str]] = re.compile(r"\d+(?:\.\d+)?")
 
 
@@ -948,8 +995,12 @@ def _try_extract_discrete_count_anchor(
     if _count_quantity_tokens(statement, padded_lower) != 1:
         return None
 
-    # Narrowness #1 + #5 — shape + counted-noun lemma.
-    extract_re = _extract_discrete_count_re_for(observed_nouns)
+    # Narrowness #1 — shape. ADR-0192: the counted-noun slot is open
+    # (adjective* + multi-word head) rather than gated on the closed
+    # observed_counted_nouns set; the other narrowness layers above plus
+    # the downstream ADR-0191 completeness guard / round-trip / branch
+    # disagreement hold wrong=0 without the curated noun list.
+    extract_re = _extract_discrete_count_re_open(observed_nouns)
     m = extract_re.match(statement.strip())
     if m is None:
         return None

tests/test_adr_0163_d2_discrete_count_injection.py

@@ -162,9 +162,17 @@ class TestExtractionRefusal:
         # Two digit runs — v1 admits exactly one count.
         assert _try_extract("He has 2 horses, 5 dogs.") is None
 
-    def test_unobserved_counted_noun_refused(self) -> None:
-        # 'widgets' is not in the spec's observed_counted_nouns.
-        assert _try_extract("Sam has 5 widgets.") is None
+    def test_unobserved_counted_noun_now_admits(self) -> None:
+        # ADR-0192 — the counted-noun slot is OPEN: an unobserved noun
+        # ('widgets', not in the spec's observed_counted_nouns) admits
+        # under the simple possession shape. The other narrowness layers
+        # (subject/verb/count/clause) and the downstream ADR-0191
+        # completeness guard + round-trip hold wrong=0, not the noun list.
+        result = _try_extract("Sam has 5 widgets.")
+        assert result is not None
+        assert result["counted_noun"].lower() == "widgets"
+        assert result["count_token"] == "5"
+        assert result["anchor_kind"] == "possession"
 
     def test_non_possession_non_acquisition_verb_refused(self) -> None:
         # Post-W2 (ADR-0170): possession verbs (has/have/had) AND

tests/test_discrete_count_open_noun_class.py

@@ -0,0 +1,62 @@
+"""ADR-0192 — open the discrete_count counted-noun class.
+
+The discrete_count matcher gated the counted noun against a CLOSED ratified
+set (``observed_counted_nouns``): "Betty has 24 marbles" matched only
+because "marbles" was ratified, while "Randy has 60 mango trees" / "Sam has
+12 red apples" emitted no anchor purely because the noun was unseen.
+
+This opens the single-anchor possession/acquisition path to an open
+noun-phrase (adjective* + multi-word head), keeping every other narrowness
+layer (proper-noun subject, possession/acquisition verb whitelist, single
+numeric token, no clause-split). Wrong=0 is held downstream by the ADR-0191
+completeness guard + round-trip + branch-disagreement — not by the curated
+noun list.
+"""
+from __future__ import annotations
+
+import pytest
+
+from generate.recognizer_match import match as rmatch
+from generate.math_candidate_graph import _load_ratified_registry_or_empty
+from generate.recognizer_anchor_inject import inject_from_match
+
+_REG = _load_ratified_registry_or_empty()
+
+
+def _anchors(sentence: str) -> int:
+    m = rmatch(sentence, _REG, prior_subject=None) if _REG else None
+    return len(m.parsed_anchors) if m is not None else -1
+
+
+# --- Now-extractable open-vocabulary possession/acquisition statements ----
+@pytest.mark.parametrize("sentence", [
+    "Randy has 60 mango trees.",            # multi-word head
+    "Randy has 60 trees on his farm.",      # single head + trailing PP
+    "Randy has 60 mango trees on his farm.",# both
+    "Sam has 12 red apples.",               # adjective + head
+    "Tom bought 5 green bottles.",          # acquisition + adjective
+])
+def test_open_noun_now_extracts(sentence: str) -> None:
+    assert _anchors(sentence) == 1, f"expected one anchor for {sentence!r}"
+
+
+def test_baseline_single_word_still_works() -> None:
+    """The previously-working closed-set case is unchanged."""
+    assert _anchors("Betty has 24 marbles.") == 1
+
+
+# --- Noun phrase must NOT swallow the trailing prepositional phrase -------
+def test_noun_phrase_stops_before_preposition() -> None:
+    m = rmatch("Randy has 60 mango trees on his farm.", _REG, prior_subject=None)
+    assert m is not None and m.parsed_anchors
+    assert m.parsed_anchors[0]["counted_noun"].lower() == "mango trees"
+
+
+# --- wrong=0 guards: shapes that MUST still refuse (no anchor) -------------
+@pytest.mark.parametrize("sentence", [
+    "Julie is reading a 120-page book.",          # verb not possession/acquisition
+    "Randy has many apples.",                      # indefinite quantifier, no count
+    "Randy has 60 apples and 30 oranges.",         # clause/enumeration split
+])
+def test_dangerous_shapes_still_refuse(sentence: str) -> None:
+    assert _anchors(sentence) == 0, f"expected no anchor for {sentence!r}"