diff --git a/CHANGELOG.md b/CHANGELOG.md
index 84bea04f..ae69f27a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,40 @@ All notable changes to Massing. Releases are signed, auto-updating desktop build
 (Windows / macOS / Linux); the updater always serves the latest. Format loosely follows
 [Keep a Changelog](https://keepachangelog.com/).
 
+## Unreleased — the app.ts figure was wrong in four places, and the gate that was supposed to stop that reached one of them
+
+**Corrects `docs/roadmap.md`, and widens `services/api/test_claude_md_gates.py` to the class rather
+than the instance.**
+
+The R39-DECOMP-VIEWER entry carried **three different figures for one file**: a header saying
+`app.ts` is "5,160 → **3,311**, a 36% cut", a shipped block saying "5,160 → **2,944**, a 43% cut",
+and the ratchet pinning **2,508**. A fourth sat elsewhere in the file — "5,064 → 3,444 → and thirteen
+slices since". The header also claimed "seven slices shipped" directly above a list naming eleven.
+
+**The two baselines were both real, and neither was labelled.** `CLAUDE.md` measures from 5,064 —
+the file before the first extraction commit of 2026-08-06 — while the roadmap measured from 5,160, a
+**same-day peak** partway through it: the file grew before it shrank. Nothing on the page could tell
+a reader the two were counted from different points. The baseline is now stated, matches `CLAUDE.md`,
+and the current figure is the ratchet's own count.
+
+**The gate is the real subject here.** The previous release added a check binding `CLAUDE.md`'s
+narrative `app.ts` figure to the `test_file_sizes.py` pin, on the reasoning that *prose is a copy of
+a gated value and a copy is what drifts*. That reasoning was right and the implementation was too
+narrow: it gated **one copy of a number that lives in four places**, and the other three were already
+wrong when it shipped. *Gating the instance you happened to be looking at is not gating the class.*
+The check now scans every doc that carries the baseline-anchored form.
+
+**It immediately earned that.** Having corrected the two figures I found by reading, the widened
+check failed on a **third** I had missed entirely.
+
+**And mutation testing found a defect in the widened check itself.** Its first draft used a summed
+floor — "at least 2 figures across the docs" — and reported "3 figures across 2 docs" while **all
+three came from the roadmap**: `CLAUDE.md` words it "5,064 lines to", so the arrow pattern never
+matched it and its contribution was zero. A mutation that reformatted only the roadmap took the
+count to 0 and exposed it. That is precisely the per-doc vacuity failure this same file's citation
+ratchet was built to prevent, repeated one screen further down. The floor is now **per-doc**, and
+`CLAUDE.md` is covered by name rather than by assumption.
+
 ## Unreleased — SCALE-SEAM (104): coverage maps
 
 Extracts `spineTraceability` and `scopeRegister` from `client.ts` (**572 → 554**) into
diff --git a/docs/roadmap.md b/docs/roadmap.md
index dd5d4d46..0281d072 100644
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@@ -1129,7 +1129,7 @@ four phantom entries — one of them a bare `SCALE-SEAM` with its `㉘` silently
 exact failure `roadmapLanes.test.ts` documents in its `MARKS` note. The gates caught all four.)*
 
 * **Not R39-DECOMP-VIEWER ③.** It ranked second last time on a size-ceiling argument that had
-  already gone false. `app.ts` is decomposing steadily (5,064 → 3,444 → and thirteen slices since), the ratchet
+  already gone false. `app.ts` is decomposing steadily (5,064 → 2,508, slices still landing), the ratchet
   is pinned, and it moves on its own whenever a feature pushes it. It does not need a sprint; it
   needs to keep being interleaved. **Re-measure the ceiling before ever promoting it again** — that
   is the specific error row 2 made.
@@ -2546,8 +2546,10 @@ re-open): the converter build stage moved to the supported Node LTS with a pinne
 refusal (`services/api/src/aec_api/main.py`), and full-history checkout for the secret-scan job.
 
 
-- 🟡 **R39-DECOMP-VIEWER ③** *(L, Lane E — **seven slices shipped; `app.ts` is 5,160 → 3,311, a 36%
-  cut. The `builders` map is entirely gone.** The paragraph below saying the extraction "is NOT begun"
+- 🟡 **R39-DECOMP-VIEWER ③** *(L, Lane E — **`app.ts` is 5,064 → 2,508, a 50% cut. The `builders`
+  map is entirely gone.** Slice count deliberately not restated here: `services/api/test_file_sizes.py`
+  is the record, and a count repeated in prose is a copy that drifts — this header said "seven slices"
+  while the list immediately below it named eleven. The paragraph below saying the extraction "is NOT begun"
   was true on 2026-08-06 and stayed on the page until 2026-08-17, through six shipped slices — a
   roadmap entry describing work as un-started while the work is being done is worse than a missing
   entry, because it sends the next reader to re-derive a plan that was already executed. Kept only
@@ -2559,10 +2561,18 @@ refusal (`services/api/src/aec_api/main.py`), and full-history checkout for the
   ⑤ project-browser panel (216) · ⑥ `loadProjectModel` (37) · ⑦ **drawings & sheets (142, v0.3.978)** ·
   ⑨ **fabrication detail (65)** · ⑩ **MEP / fire / life safety (169)** — both v0.3.981 ·
   ⑫ **envelope & free-form geometry (75, v0.3.982)** · ⑬ **model federation & version compare
-  (88, v0.3.1043)**. `app.ts` 5,160 → **2,944**, a **43% cut**.
+  (88, v0.3.1043)**. `app.ts` 5,064 → **2,508**, a **50% cut**.
   Each ratcheted `services/api/test_file_sizes.py` down, never reset. `services/api/test_file_sizes.py`
   carries the per-slice history; that comment, not this list, is the record.
 
+  *(**Three figures for one file, two of them stale — corrected 2026-09-05.** This block said
+  `5,160 → 2,944` and the header said `5,160 → 3,311`, while the ratchet pinned **2,508**. The two
+  BASELINES also disagreed with `CLAUDE.md`, which says 5,064, and both were real: 5,064 is the file
+  before the first extraction commit of 2026-08-06, and **5,160 was a same-day peak** partway through
+  it — the file grew before it shrank. Neither number was ever labelled, so nothing could tell a
+  reader they were measuring from different points. Baseline is now stated and matches `CLAUDE.md`;
+  the current figure is the ratchet's own count, which is the only one under a gate.)*
+
   **"never reset" was false for two days, and the way it was found is the transferable part.** Slices
   ⑭⑮⑯ walked the pin 2_865 → 2_757 → 2_630 → 2_571 on 2026-08-27, and a REL-4 portal commit to the
   same shared file later that afternoon carried the pre-⑭ line back in — value *and* comment trail, so
diff --git a/services/api/test_claude_md_gates.py b/services/api/test_claude_md_gates.py
index 4fce2ed8..5c1cb2c0 100644
--- a/services/api/test_claude_md_gates.py
+++ b/services/api/test_claude_md_gates.py
@@ -289,6 +289,53 @@ def resolve(name):
 # Both halves must be FOUND, not just agree. If either pattern stops matching — the sentence is
 # reworded, the ratchet key is renamed — this check would otherwise pass on two Nones, which is the
 # vacuous-green failure this file's own header calls worse than no gate.
+# The SAME figure lives in `docs/roadmap.md` twice, and the gate above did not reach it. Added
+# 2026-09-05, one PR after that gate: the roadmap's R39-DECOMP-VIEWER entry carried THREE figures for
+# `app.ts` — a header saying "5,160 -> 3,311", a shipped block saying "5,160 -> 2,944", and the truth,
+# 2,508. It also used a DIFFERENT baseline from CLAUDE.md (5,160 vs 5,064) with neither labelled, so
+# nothing could tell a reader they were measured from different points; both were real, 5,064 being the
+# file before the first extraction commit and 5,160 a same-day peak partway through it.
+#
+# The lesson is about the gate, not the numbers. The check above was written to stop a narrative copy of
+# a pinned value from drifting, and it gated ONE copy of a number that lives in three places. **Gating
+# the instance you happened to be looking at is not gating the class.** So this scans every doc for the
+# baseline-anchored form and requires each to equal the pin — which also holds the BASELINE steady,
+# since a figure measured from somewhere else no longer matches the pattern at all.
+#
+# Historical mentions are deliberately still matchable only by their own baseline: the corrected text
+# quotes "5,160 -> 2,944" as the thing that was wrong, and that does not match this pattern, so the
+# record of the error survives without failing the build.
+APP_TS_BASELINE = "5,064"
+_arrow = re.compile(re.escape(APP_TS_BASELINE) + r"\s*(?:->|\u2192)\s*\*{0,2}([\d,]+)")
+
+#: Docs carrying the ARROW form, and the minimum each must contribute. PER-DOC, never summed — the
+#: first draft of this check used a summed floor of 2 and reported "3 figures across 2 docs" while
+#: **all three came from the roadmap**: CLAUDE.md words it "5,064 lines to", so the arrow pattern
+#: never matched it and its contribution was ZERO. A mutation that reformatted only the roadmap took
+#: the count to 0 and revealed it. That is the exact failure this file's own citation ratchet was
+#: built to avoid, repeated one screen further down — **a summed floor lets one doc hold the number
+#: up while another silently contributes nothing.** CLAUDE.md is not listed here because the two
+#: checks below cover its wording directly; a doc is either in this map or covered by name, never
+#: assumed.
+ARROW_DOCS = {"docs/roadmap.md": 3}
+
+_bad, _thin = [], []
+for _doc, _min in ARROW_DOCS.items():
+    with open(os.path.join(ROOT, _doc), encoding="utf-8") as fh:
+        _hits = _arrow.findall(fh.read())
+    if len(_hits) < _min:
+        _thin.append(f"{_doc}={len(_hits)} (min {_min})")
+    for _h in _hits:
+        if _pinned and int(_h.replace(",", "")) != int(_pinned.group(1).replace("_", "")):
+            _bad.append(f"{_doc}: {APP_TS_BASELINE} -> {_h}")
+
+check(
+    "every doc that quotes the app.ts figure quotes the pinned one",
+    not _bad and not _thin,
+    (f"below the floor: {', '.join(_thin)}; " if _thin else "")
+    + (f"DISAGREE: {', '.join(_bad)}" if _bad else "all figures match the pin"),
+)
+
 check(
     "CLAUDE.md's app.ts sentence and the ratchet pin are both readable",
     bool(_stated) and bool(_pinned),