fix(shapes): the semantic arm proposes at the write-path floor, not a private 0.8 (#4208)

0.8 was sized for proposals nobody reads; every semantic proposal is read
by a judge before it is confirmed. Measured, it proposed 0 of 150 while
judged instances of a canon score 0.68-0.71 - the write-path hint's own
floor asks the same question of the same documents at 0.68. The arm now
uses that floor, scans 8 hits instead of 3 (a true instance ranked 4th
behind snippets of other language families), and the proposer version
bumps to 5 so rows examined under the old floor are read again.

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

src/scribe/services/shape_ledger.py

@@ -1516,24 +1516,32 @@ _DERIVE_MIN_NAME_CSS = 2
 # Semantic checks per repo per refresh — an embedding each (local fastembed),
 # bounded so a 4,000-row ledger is worked through over refreshes, not in one.
 _SEMANTIC_CAP = 150
-# The semantic basis's own floor — stricter than the write-path arm's, because
+# The semantic arm has NO floor of its own: it uses the write-path surface's,
-# a code BODY against a prose-forward snippet document scores 0.68–0.75 for
+# because it asks the write-path hint's question — "which canon does this body
-# "both are about migrations"; first live run paired every alembic
+# resemble?" — of the same snippet documents (#4208).
-# upgrade()/downgrade() with an unrelated canon at exactly that band.
-_SEMANTIC_FLOOR = 0.8
-# How many above-floor hits the semantic arm asks for.
-_SEMANTIC_LIMIT = 3
-# A MISS FROM THIS ARM IS NOT EVIDENCE, and nothing may read it as one (#4208).
 #
+# It had one, 0.8, set after the first live run paired alembic
+# upgrade()/downgrade() with unrelated canons at 0.68–0.75. That floor was
+# sized for a proposal nobody reads, and nobody-reads is not how proposals are
+# consumed: a judge reviews each semantic one (confirm_proposals refuses an
+# unnamed batch, and the tool says "semantic deserves a look"). Measured on
+# 2026-09-22 it proposed 0 of 150, while judged instances of the service-unit
+# canon score 0.68–0.71 against it. A wrong proposal costs the judge a look; a
+# floor that yields nothing costs the signal. Noise in this band is expected —
+# the score rides on the proposal so the judge can weigh it.
+#
+# How many hits the arm scans for an ALLOWED canon. Wider than the hint's
+# budget because disallowed snippets (other language families) share the
+# ranking: one measured true instance sat 4th, behind three it could not use.
+_SEMANTIC_LIMIT = 8
+# A MISS FROM THIS ARM IS NOT EVIDENCE, and nothing may read it as one (#4208).
 # #4208 briefly stored "compared, nothing cleared the floor" and let it
-# withdraw divergence prompts. Measured live on 2026-09-22, with the arm's own
+# withdraw divergence prompts. It silenced real divergences exactly as it
-# query shape: judged instances of the service-unit canon (#2860) score 0.68–
+# silenced helpers: most judged instances of a canon score below any floor
-# 0.71 against it at best, and most fall below 0.66; helpers beside it score
+# that keeps noise out (0.66 and under, in the measurement above), and helpers
-# 0.66–0.75 against snippets they have nothing to do with. Nothing reaches
+# score 0.66–0.75 against snippets they have nothing to do with. A code body
-# 0.8, so the arm answers "no canon" for almost every body, true instance or
+# against a prose-forward snippet document measures the wrong field (#2518);
-# not — a real divergence was silenced exactly as a helper was. A code body
+# no floor separates the two bands. The arm PROPOSES on a hit; its silence
-# against a prose-forward snippet document measures the wrong field (#2518),
-# and no floor separates the two bands. The arm PROPOSES on a hit; its silence
 # says nothing.
 
 
@@ -1561,7 +1569,9 @@ def _semantic_priority(row) -> tuple:
 # v4: the semantic arm recorded its misses too (#4208). Retired — a miss is
 # not evidence (see _SEMANTIC_LIMIT) — but the proposals themselves did not
 # change, so no re-examination is owed; a stale miss basis is inert.
-_PROPOSER_VERSION = 4
+# v5: the semantic arm's floor dropped from 0.8 to the write-path floor (#4208),
+# so rows it examined and found nothing for must be read once more.
+_PROPOSER_VERSION = 5
 # Signature resemblance floor, name blanked (difflib ratio) — and a length
 # floor, because `def NAME():` resembles `def NAME(x):` at 0.95 while saying
 # nothing; a family shape has parameters to resemble.
@@ -1810,7 +1820,7 @@ async def _semantic_canon(
     query = concept_query(body) or body
     hits = await semantic_search_notes(
         user_id, query, limit=_SEMANTIC_LIMIT,
-        threshold=max(WRITEPATH_DEFAULT_THRESHOLD, _SEMANTIC_FLOOR),
+        threshold=WRITEPATH_DEFAULT_THRESHOLD,
         note_type="snippet", scope="browse",
     )
     for score, note in hits:

tests/test_divergence_meaning_gate.py

@@ -11,9 +11,11 @@ do. A signature does not carry a job.
 the floor" and let it withdraw the prompt. Measured live on 2026-09-22 it
 cannot discriminate. Judged instances of the service-unit canon score 0.68–
 0.71 against it at best, most below 0.66; helpers score 0.66–0.75 against
-snippets unrelated to them. Nothing reaches the 0.8 floor, so the "miss"
+snippets unrelated to them. The then-floor of 0.8 was reached by nothing, so
-fires for nearly every body — the real divergence silenced exactly as the
+the "miss" fired for nearly every body — the real divergence silenced exactly
-helper was. The gate was removed; the prompt asks and the judge answers.
+as the helper was. The gate was removed; the prompt asks and the judge
+answers. The same measurement moved the arm's floor down to the write-path
+hint's: 0.8 was sized for proposals nobody reads, and it proposed 0 of 150.
 
 These tests pin what stayed: the arm proposes on a hit and says nothing on a
 miss, the divergence check does not consult it, and the capped pass reads new
@@ -92,6 +94,18 @@ async def test_no_allowed_canon_spends_no_search() -> None:
     mock.assert_not_awaited()
 
 
+async def test_the_arm_proposes_at_the_write_path_floor() -> None:
+    """It asks the write-path hint's question of the same documents, so it
+    uses that surface's floor — not a stricter private one that yields no
+    proposals for the judge to weigh (#4208)."""
+    from scribe.services.plugin_context import WRITEPATH_DEFAULT_THRESHOLD
+
+    mock = _hits()
+    with _patch(mock):
+        await _semantic_canon(1, BODY, {CANON})
+    assert mock.await_args.kwargs["threshold"] == WRITEPATH_DEFAULT_THRESHOLD
+
+
 # ── ...and its silence reaches nothing ───────────────────────────────────
 
 

tests/test_integration_shape_classify.py

@@ -897,7 +897,7 @@ async def test_a_semantic_miss_does_not_silence_the_prompt(seeded):
     `confirmDanger` beside an async confirm canon is structurally identical to
     a registry helper beside an async service canon, and the arm's miss was
     meant to tell them apart. Measured live, true instances of a canon rarely
-    clear the arm's 0.8 floor either, so a miss fires on both — silencing on
+    clear the arm's floor either, so a miss fires on both — silencing on
     it silenced #2793's own case. Here the arm has read the body, found
     nothing, and the prompt is still RAISED.