fix(drafter): the base rate counted a dataclass and an async service unit as comparable things (#4208)

A CORRECTION TO THIS ISSUE'S OWN PLAN, first, because the plan was wrong in
a way that would have cost the acceptance case.

#4208 recommended widening `kind` past `css | sym` to carry the form, calling
the denominator the deeper win and a migration the price. Two things:

1. THE MIGRATION IS NOT NEEDED. The issue says density "is computed from
   stored rows and `shape_form` only runs on read". `canon_density` does
   `select(CodeShape)...scalars().all()` and filters to the directory in
   PYTHON; `signature` is a stored column; `canon_form(siblings, ...)` on the
   next line already derives form from those rows. Bucketing the denominator
   costs a list comprehension.

2. WIDENING `kind` WOULD BREAK MILESTONE #2793's ACCEPTANCE CASE. If `kind`
   separated `fn` from `async-fn`, density would bucket them apart — which
   does silence #4204's four `def` helpers, as claimed. Apply it to a
   hand-rolled SYNC `confirmDanger` in a directory where an async confirm
   helper is canon: the canon leaves the candidate's denominator, nothing
   dominates, no flag. That is the flag the milestone exists to produce, and
   it is the same inversion the first form gate made, one layer down.

So option 1 is not expensive-but-right, it is wrong; and option 2 (compare
meaning) is not the interim, it is the only lever. `confirmDanger` beside an
async confirm helper and `is_registered` beside a service unit are identical
at every structural level — same family, same form contradiction, same
density. They differ only in whether the candidate does the canon's JOB, and
no signature carries that.

WHAT THIS SHIPS is the part that is right and provable: an honest
denominator. `comparable_siblings` narrows the count to rows whose family
does not contradict the candidate's, using `families_conflict` — the same
predicate the gate uses, so the count and the verdict cannot drift into
disagreeing about what comparable means. "372 judged siblings" stops counting
a dataclass, a constant and an async service unit as three comparable things.

NARROWED BY FAMILY, NOT FORM, for the reason above: `fn` beside `async-fn`
stays a fair question. An unreadable sibling STAYS COUNTED — dropping it
would shrink `judged`, raise the share, and fire the check more on the
directories it can read least. Every unknown-form decision in this module
goes that way.

Density is now per candidate rather than per kind, cached on (kind, form).

TWO BUGS THIS CHANGE HAD, both caught before CI and both pinned:

- I passed a FAMILY where `families_conflict` reads a FORM, so
  `shape_family("callable")` returned "" and the narrowing was a silent
  no-op that still read as applied. The regression guard deliberately uses a
  callable: `type` is both a form and a family name, so testing with it
  proves nothing and the bug hides.
- My new `_Row` in the test file shadowed the one already there — same
  fields, different `snippet_id` default — silently breaking three passing
  `canon_form` tests. Reused the existing class. A duplicate definition
  quietly changing a neighbour's meaning is this file's own subject.

The residue test's docstring said separating the four needs "widen `kind` or
a comparison of meaning". Corrected: they are not alternatives.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01821k5B3Ysecp9fNYs92Kuy

src/scribe/services/shape_ledger.py

@@ -2135,6 +2135,35 @@ _DENSITY_MIN_JUDGED = 3
 _DENSITY_SHARE = 0.6
 
 
+def comparable_siblings(rows: Iterable, form: str) -> list:
+    """The siblings a candidate of ``form`` can honestly be counted against.
+
+    Excludes only a KNOWN family contradiction, via `families_conflict` — the
+    same predicate the divergence gate uses, so the denominator and the gate
+    cannot drift into disagreeing about what "comparable" means.
+
+    `form` is a FORM (`fn`, `async-fn`, `type`, …), never a family:
+    `families_conflict` coarsens both sides itself, and handing it a family
+    makes `shape_family("callable")` return "" so nothing is excluded — a
+    narrowing that silently becomes a no-op while still reading as applied.
+
+    AN UNREADABLE SIBLING STAYS COUNTED, and that direction is the point.
+    Dropping it would shrink `judged`, raise the dominant canon's share, and
+    make the check fire MORE on the directories it can read least. Every
+    unknown-form decision in this module goes the same way: quieter, never
+    more confident.
+    """
+    if not form:
+        return list(rows)
+    return [
+        r for r in rows
+        if not families_conflict(
+            shape_form(getattr(r, "signature", "") or "", getattr(r, "kind", "sym")),
+            form,
+        )
+    ]
+
+
 def dominant_canon(rows: Iterable[CodeShape]) -> tuple[int, int, int] | None:
     """(snippet_id, its_count, judged_count) when one canon dominates these
     sibling rows (same directory + kind), else None."""
@@ -2156,9 +2185,37 @@ def _dir_of(path: str) -> str:
     return path.rsplit("/", 1)[0] if "/" in path else ""
 
 
-async def canon_density(project_id: int, path: str, kind: str) -> tuple[int, int, int] | None:
+async def canon_density(
+    project_id: int, path: str, kind: str, form: str = ""
+) -> tuple[int, int, int, str] | None:
     """The dominant canon for the directory ``path`` sits in, for ``kind`` —
-    the write-time question "is this a canon-dense place?"."""
+    the write-time question "is this a canon-dense place?".
+
+    ``form`` is the candidate's own FORM — `fn`, `async-fn`, `type`, … as
+    `shape_form` returns it, NOT a family. `families_conflict` coarsens both
+    sides itself, and handing it a family makes `shape_family("callable")`
+    return "" so nothing is ever excluded: the narrowing silently becomes a
+    no-op that still reads as applied. It narrows the DENOMINATOR
+    (#4208). Without it the base rate was computed over every code symbol in
+    the directory as one bucket: "372 judged siblings" counted a dataclass, a
+    CSS-less constant and an async service unit as three comparable things,
+    and the share that came out of that was a statement about a population
+    nobody had asked a question about.
+
+    EXCLUDES ONLY A KNOWN CONFLICT, using `families_conflict` — the same
+    predicate the divergence gate uses, so the two cannot drift apart. A
+    sibling whose form is unreadable STAYS COUNTED. That direction is
+    deliberate and it is the one that matters: dropping unknown rows would
+    shrink `judged`, raise the share, and make the check fire MORE on exactly
+    the directories it can read least. Every other unknown-form decision in
+    this module goes the same way — quieter, never more confident.
+
+    NOT narrowed to the candidate's exact form, for the reason `shape_family`
+    gives at length: `fn` beside `async-fn` is the acceptance case of
+    milestone #2793, not noise. Bucketing the denominator by form would take
+    the async canon out of a sync candidate's count and silence that flag —
+    the same inversion the first form gate made, one layer down.
+    """
     directory = _dir_of(path)
     async with async_session() as session:
         rows = (
@@ -2173,6 +2230,7 @@ async def canon_density(project_id: int, path: str, kind: str) -> tuple[int, int
             )
         ).scalars().all()
     siblings = [r for r in rows if _dir_of(r.path) == directory]
+    siblings = comparable_siblings(siblings, form)
     dom = dominant_canon(siblings)
     if dom is None:
         return None
@@ -2195,10 +2253,28 @@ async def write_time_divergence(
     edit. Returns [{symbol, kind, canon_snippet_id, instances, judged}]."""
     just_stamped = {(s["symbol"], s["kind"]): s["snippet_id"] for s in stamped}
     out: list[dict] = []
-    kinds = {k for k, _n in shapes}
+    if not shapes:
-    density = {k: await canon_density(project_id, path, k) for k in kinds}
-    if not any(density.values()):
         return out
+    # DENSITY IS NOW PER CANDIDATE, not per kind (#4208): the denominator
+    # excludes siblings whose family contradicts what is being written, so it
+    # cannot be computed until the candidate's own form is known. Cached on
+    # (kind, family) — a write names a handful of shapes and they collapse to
+    # one or two buckets, so this is the same one-or-two queries as before.
+    #
+    # The cost is that the row load below no longer sits behind an early exit
+    # on "nothing is dense here". That is one indexed lookup on
+    # (project_id, path), and it has to happen first regardless: the
+    # candidate's signature comes from its stored row when the payload does
+    # not carry one.
+    _density: dict[tuple[str, str], tuple[int, int, int, str] | None] = {}
+
+    async def density_for(kind: str, form: str):
+        # Keyed and passed as a FORM, not a family — see `canon_density`.
+        key = (kind, form)
+        if key not in _density:
+            _density[key] = await canon_density(project_id, path, kind, form)
+        return _density[key]
+
     async with async_session() as session:
         rows = (
             await session.execute(
@@ -2211,13 +2287,19 @@ async def write_time_divergence(
         ).scalars().all()
     by_key = {(r.symbol, r.kind): r for r in rows}
     for kind, name in shapes:
-        dom = density.get(kind)
+        row = by_key.get((name, kind))
+        # The candidate's own form, read before anything is counted. Signature
+        # from the payload first — a shape being written now may have no row
+        # yet — falling back to the stored row's.
+        mine = shape_form(
+            signature_in(code, name, kind) or getattr(row, "signature", "") or "", kind
+        )
+        dom = await density_for(kind, mine)
         if not dom:
             continue
         sid, n, judged, cform = dom
         if just_stamped.get((name, kind)) == sid:
             continue
-        row = by_key.get((name, kind))
         if row is not None and (
             row.status != "unclassified" or row.proposed_snippet_id == sid
         ):
@@ -2231,12 +2313,10 @@ async def write_time_divergence(
         # divergence prompt is ABOUT a mismatch, so requiring the candidate to
         # match would silence the check precisely where it belongs.
         #
-        # Signature from the payload first — a shape being written now may
+        # `mine` was read above, before the denominator was counted — the
-        # have no row yet — and an unreadable one produces a fair question
+        # same value serves both, and they must not be able to disagree.
-        # rather than a guess, because `families_conflict` needs both sides.
+        # An unreadable signature produces a fair question rather than a
-        mine = shape_form(
+        # guess, because `families_conflict` needs both sides.
-            signature_in(code, name, kind) or getattr(row, "signature", "") or "", kind
-        )
         if families_conflict(mine, cform):
             continue
         out.append({"symbol": name, "kind": kind, "canon_snippet_id": sid,

tests/test_shape_form_gate.py

@@ -26,8 +26,8 @@ from __future__ import annotations
 import pytest
 
 from scribe.services.shape_ledger import (
-    FORM_UNKNOWN, canon_form, families_conflict, forms_agree, shape_family,
+    FORM_UNKNOWN, canon_form, comparable_siblings, dominant_canon,
-    shape_form, signature_in,
+    families_conflict, forms_agree, shape_family, shape_form, signature_in,
 )
 
 
@@ -103,8 +103,17 @@ def test_how_many_of_the_five_the_divergence_gate_actually_silences() -> None:
     That is not a shortcoming of the gate, it is the limit of the signature:
     at this level those four are indistinguishable from #2793's acceptance
     case, where a sync `confirmDanger` beside an async confirm helper SHOULD
-    be flagged. Separating them needs #4204 option 2 (widen `kind`) or a
+    be flagged.
-    comparison of meaning."""
+
+    CORRECTED 2026-09-21 (#4208). This used to say separating them needs
+    "widen `kind` or a comparison of meaning", offering the two as
+    alternatives. Widening `kind` does not separate them: it buys the silence
+    by bucketing `fn` apart from `async-fn`, which takes the async canon out
+    of a sync candidate's denominator and silences #2793's acceptance case by
+    the identical mechanism, one layer down. Whatever separates these four
+    has to distinguish a helper that does the canon's JOB from one that does
+    not, and no signature carries that. A comparison of meaning is the only
+    lever, not one of two."""
     canon = shape_form("async def create_note(user_id: int, ...):", "sym")
     silenced = {
         sig: families_conflict(shape_form(sig, "sym"), canon)
@@ -323,3 +332,112 @@ def test_the_resemblance_floor_is_above_the_retrieval_floors() -> None:
     from scribe.services.shape_ledger import _RESEMBLE_MIN
 
     assert _RESEMBLE_MIN >= 0.80
+
+
+# ── The denominator, not the gate (#4208) ─────────────────────────────────
+#
+# `dominant_canon` is a base rate, and a base rate is only a statement about
+# something if its denominator is a population somebody asked a question
+# about. It counted every code symbol in a directory as one bucket: a frozen
+# dataclass, a module constant and an async service unit were three
+# comparable things, and "372 judged siblings" was the authority the
+# divergence line spoke with.
+#
+# `comparable_siblings` narrows it using the SAME predicate as the gate, so
+# the count and the verdict cannot drift into disagreeing about what
+# comparable means.
+
+
+# `_Row` above is reused rather than redefined. A second class of the same
+# name here shadowed the first — same fields, different `snippet_id` default —
+# and silently broke three `canon_form` tests that had been passing, which is
+# a neater demonstration of this file's subject than anything it asserts.
+SERVICE = "async def get_note(user_id: int, note_id: int) -> Note | None:"
+HELPER = "def is_registered(source: str) -> bool:"
+
+
+def test_a_type_is_not_counted_against_a_directory_of_callables() -> None:
+    rows = [_Row(SERVICE) for _ in range(6)] + [_Row("class Point:", snippet_id=9)]
+    assert len(comparable_siblings(rows, shape_form("class Point:"))) == 1
+
+
+def test_a_callable_is_not_counted_against_a_dataclass() -> None:
+    """The honest-denominator half, and the one that changes reported numbers:
+    `judged` stops overstating how much of the directory was ever comparable."""
+    rows = [_Row(SERVICE) for _ in range(6)] + [_Row("class Point:", snippet_id=9)]
+    assert len(comparable_siblings(rows, shape_form(HELPER))) == 6
+
+
+def test_the_async_canon_stays_in_a_sync_candidates_denominator() -> None:
+    """THE ACCEPTANCE CASE OF #2793, and the reason this narrows by family
+    rather than by form.
+
+    A hand-rolled sync `confirmDanger` in a directory where an async confirm
+    helper is canon must still be flagged. Bucketing the denominator by exact
+    form — which is what widening `kind` to carry `fn` vs `async-fn` amounts
+    to — would take the canon out of this count and silence it.
+    """
+    rows = [_Row("async def confirmDanger(message: str) -> bool:") for _ in range(5)]
+    kept = comparable_siblings(rows, shape_form("function confirmDanger(message) {"))
+    assert len(kept) == 5
+    assert dominant_canon(kept) is not None
+
+
+def test_an_unreadable_sibling_stays_counted() -> None:
+    """Quieter, never louder. Dropping unknown rows shrinks `judged`, raises
+    the dominant canon's share, and fires the check MORE on exactly the
+    directories it can read least."""
+    rows = [_Row(SERVICE) for _ in range(4)] + [_Row("")]
+    assert len(comparable_siblings(rows, shape_form(SERVICE))) == 5
+
+
+def test_an_unreadable_candidate_narrows_nothing() -> None:
+    """The other side of the same discipline: a candidate whose own signature
+    says nothing gets the full denominator, not a guessed one."""
+    rows = [_Row(SERVICE), _Row("class Point:")]
+    assert comparable_siblings(rows, FORM_UNKNOWN) == rows
+
+
+def test_a_family_passed_where_a_form_belongs_would_be_a_silent_no_op() -> None:
+    """A REGRESSION GUARD ON A BUG THIS CHANGE ACTUALLY HAD.
+
+    `families_conflict` coarsens both sides itself, so handing it a family
+    makes `shape_family("callable")` return "" and the whole narrowing becomes
+    a no-op — while every call site still reads as though it applied. Pinned
+    because the wrong value is the right TYPE and the failure is silent.
+    """
+    # A CALLABLE candidate, deliberately: `fn`'s family is `callable`, a word
+    # that is not itself a form, so `shape_family("callable")` is "" and the
+    # exclusion never fires. Picking `type` here would prove nothing — `type`
+    # is both a form and a family name, so passing the family still narrows
+    # and the bug hides. That near-miss is why this test exists at all.
+    rows = [_Row(SERVICE) for _ in range(6)] + [_Row("class Point:", snippet_id=9)]
+    form = shape_form("def helper(x) -> bool:")
+    assert shape_family(form) != form, "this test needs a form whose family differs"
+    by_form = comparable_siblings(rows, form)
+    by_family = comparable_siblings(rows, shape_family(form))
+    assert len(by_form) == 6, "the dataclass is not comparable to a function"
+    assert len(by_family) == len(rows), "a family narrows nothing — that is the bug"
+    assert by_form != by_family
+
+
+def test_the_four_survivors_still_prompt() -> None:
+    """HONEST ACCOUNTING, matching the sibling test above.
+
+    The narrowed denominator does not silence #4204's four `def` helpers, and
+    was never going to: they are callables, the canon is a callable, so
+    nothing is excluded from their count. What changes is that the count is
+    now over comparable things. Asserted so the claim cannot quietly rot into
+    "this fixed it".
+    """
+    rows = [_Row(SERVICE) for _ in range(6)] + [_Row("class Point:", snippet_id=9)]
+    for sig in (
+        "def _p(source, kind, what, **kw) -> tuple[str, Point]:",
+        "def get_point(source: str) -> Point | None:",
+        "def is_registered(source: str) -> bool:",
+        "def sources_expected_to_emit() -> list[str]:",
+    ):
+        kept = comparable_siblings(rows, shape_form(sig))
+        dom = dominant_canon(kept)
+        assert dom is not None, sig
+        assert not families_conflict(shape_form(sig), canon_form(kept, dom[0])), sig