diff --git a/examples/bmesh-gear/README.md b/examples/bmesh-gear/README.md
index 7347de1..851c147 100644
--- a/examples/bmesh-gear/README.md
+++ b/examples/bmesh-gear/README.md
@@ -17,10 +17,28 @@ faces). If an op leaks geometry or a face fails to close, the math catches it.
 # Cheap correctness check (no render) — the CI check:
 blender --background --python bmesh_gear.py --
 
+# Falsifier: skip the extrude. Must exit non-zero (topology).
+blender --background --python bmesh_gear.py -- --no-extrude
+
 # Also render a still (EEVEE on a GPU host; use --engine cycles on GPU-less hosts):
 blender --background --python bmesh_gear.py -- --output gear.png
 blender --background --python bmesh_gear.py -- --output gear.png --engine cycles
 ```
 
-It exits non-zero on failure (topology mismatch or non-manifold edges). The `blender-smoke`
-workflow runs the check on Blender 5.2 LTS and 4.5 LTS.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | Topology ≠ closed form (`--no-extrude` lands here) |
+| 4 | Non-manifold edges |
+| 6 | `--output` produced no file |
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--no-extrude`.
diff --git a/examples/bmesh-gear/bmesh_gear.py b/examples/bmesh-gear/bmesh_gear.py
index 3c17888..cfcc2da 100644
--- a/examples/bmesh-gear/bmesh_gear.py
+++ b/examples/bmesh-gear/bmesh_gear.py
@@ -7,10 +7,15 @@
 verts = 2 x (4 x teeth), faces = sides + 2 caps, edges = 3 x profile — and
 that the mesh is watertight (every edge borders exactly 2 faces).
 
+``--no-extrude`` skips the face-region extrude and still runs the closed-form
+count check, so verts stay at one ring. That is the falsifier
+(``--same-axis`` in export-preset-axis).
+
 By default it runs only the correctness check (no render) — the CI smoke
 check. Pass --output to also render a still:
 
     blender --background --python bmesh_gear.py --                 # check only
+    blender --background --python bmesh_gear.py -- --no-extrude    # must fail
     blender --background --python bmesh_gear.py -- --output g.png  # + render
 """
 import bpy, bmesh, sys, os, math, argparse
@@ -39,16 +44,17 @@ def gear_profile():
     return [(r * math.cos(a), r * math.sin(a), 0.0) for a, r in coords]
 
 
-def build_gear():
+def build_gear(no_extrude=False):
     bpy.ops.wm.read_factory_settings(use_empty=True)
     me = bpy.data.meshes.new("Gear")
     bm = bmesh.new()
     try:
         verts = [bm.verts.new(co) for co in gear_profile()]
         face = bm.faces.new(verts)
-        ext = bmesh.ops.extrude_face_region(bm, geom=[face])
-        top_verts = [e for e in ext["geom"] if isinstance(e, bmesh.types.BMVert)]
-        bmesh.ops.translate(bm, verts=top_verts, vec=(0.0, 0.0, DEPTH))
+        if not no_extrude:
+            ext = bmesh.ops.extrude_face_region(bm, geom=[face])
+            top_verts = [e for e in ext["geom"] if isinstance(e, bmesh.types.BMVert)]
+            bmesh.ops.translate(bm, verts=top_verts, vec=(0.0, 0.0, DEPTH))
         bmesh.ops.recalc_face_normals(bm, faces=bm.faces)
         bm.to_mesh(me)
     finally:
@@ -193,9 +199,11 @@ def main():
     p.add_argument("--output", default=None, help="optional: render a still PNG here")
     p.add_argument("--engine", default="eevee", choices=("eevee", "cycles"),
                    help="render engine for --output (cycles for GPU-less hosts)")
+    p.add_argument("--no-extrude", action="store_true",
+                   help="skip the face-region extrude (must fail)")
     args = p.parse_args(argv)
 
-    obj = build_gear()
+    obj = build_gear(no_extrude=args.no_extrude)
     code = check(obj)
     if code:
         return code
diff --git a/examples/collision-hull-proxy/README.md b/examples/collision-hull-proxy/README.md
index 003af6e..43d7716 100644
--- a/examples/collision-hull-proxy/README.md
+++ b/examples/collision-hull-proxy/README.md
@@ -61,12 +61,32 @@ through the shell in frame.
 # Cheap correctness check (no render) — the CI check:
 blender --background --python collision_hull_proxy.py --
 
+# Falsifier: shrink hull vertices. Must exit non-zero (containment).
+blender --background --python collision_hull_proxy.py -- --shrink-hull
+
 # Also render a still (EEVEE on a GPU host; use --engine cycles on GPU-less hosts):
 blender --background --python collision_hull_proxy.py -- --output hydrant.png
 blender --background --python collision_hull_proxy.py -- --output hydrant.png --engine cycles
 ```
 
-It exits non-zero on failure (render geometry escaping a hull, inverted
-winding, a non-watertight or non-convex piece, Euler drift, or a piece over
-the 255-face budget). The `blender-smoke` workflow runs the check on
-Blender 5.2 LTS and 4.5 LTS.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | Render vertex escapes its hull (`--shrink-hull` lands here) |
+| 4 | Hull edge does not border exactly two faces |
+| 5 | Signed volume ≤ 0 (inverted winding) |
+| 6 | Hull vertex off its own face plane |
+| 7 | Euler characteristic ≠ 2 |
+| 8 | Piece over the 255-face collision budget |
+| 9 | `--output` produced no file |
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--shrink-hull`.
diff --git a/examples/collision-hull-proxy/collision_hull_proxy.py b/examples/collision-hull-proxy/collision_hull_proxy.py
index 5b27568..f085aff 100644
--- a/examples/collision-hull-proxy/collision_hull_proxy.py
+++ b/examples/collision-hull-proxy/collision_hull_proxy.py
@@ -24,10 +24,15 @@
 details (grooves) never touch the hull; proud details cost hull faces. That
 trade-off IS the collision-authoring lesson.
 
+``--shrink-hull`` scales each hull's vertices by 0.5 and still runs the
+containment plane test, so render verts escape. That is the falsifier
+(``--same-axis`` in export-preset-axis).
+
 By default it runs only the correctness check (no render) — the CI smoke
 check. Pass --output to also render a still:
 
     blender --background --python collision_hull_proxy.py --                 # check only
+    blender --background --python collision_hull_proxy.py -- --shrink-hull   # must fail
     blender --background --python collision_hull_proxy.py -- --output h.png  # + render
 """
 import bpy, bmesh, sys, os, math, argparse
@@ -485,12 +490,18 @@ def main():
     p.add_argument("--output", default=None, help="optional: render a still PNG here")
     p.add_argument("--engine", default="eevee", choices=("eevee", "cycles"),
                    help="render engine for --output (cycles for GPU-less hosts)")
+    p.add_argument("--shrink-hull", action="store_true",
+                   help="scale hull vertices by 0.5 (must fail)")
     args = p.parse_args(argv)
 
     groups = build_hydrant()
     pieces = []
     for g in groups:
         hull = build_hull(f"{g['name']}Hull", collect_points(g["cage"]))
+        if args.shrink_hull:
+            for v in hull.data.vertices:
+                v.co *= 0.5
+            hull.data.update()
         pieces.append((g["name"], hull, collect_points(g["render"])))
     code = check(pieces)
     if code:
diff --git a/examples/degenerate-bevel-weld/README.md b/examples/degenerate-bevel-weld/README.md
index 398853a..15cd9a8 100644
--- a/examples/degenerate-bevel-weld/README.md
+++ b/examples/degenerate-bevel-weld/README.md
@@ -49,11 +49,30 @@ broken state is in-frame by design.
 
 ```bash
 blender --background --python degenerate_bevel_weld.py --
+blender --background --python degenerate_bevel_weld.py -- --both-safe
 blender --background --python degenerate_bevel_weld.py -- --output bevel.png
 blender --background --python degenerate_bevel_weld.py -- --output bevel.png --engine cycles
 ```
 
-Exits non-zero on failure. The `blender-smoke` workflow runs the check on
-Blender 5.2 LTS and 4.5 LTS. The `--output` render path additionally measures
-framing against the Layer 1 band via `examples/gallery_framing.py` (exit 10
-on violation) before writing the still.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it. `10` is the shared framing helper.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | Safe bevel produced zero-area faces |
+| 4 | Degenerate bevel zero-area count ≠ closed form (`--both-safe` lands here) |
+| 5 | min_area collapse under 1e5× |
+| 6 | Coincident-position count off the closed form |
+| 7 | Safe GLB carries degenerate triangles |
+| 8 | Degenerate GLB triangle or position count drifted |
+| 9 | `--output` produced no file |
+| 10 | Gallery framing violation |
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--both-safe`.
diff --git a/examples/degenerate-bevel-weld/degenerate_bevel_weld.py b/examples/degenerate-bevel-weld/degenerate_bevel_weld.py
index cfd6d78..66288aa 100644
--- a/examples/degenerate-bevel-weld/degenerate_bevel_weld.py
+++ b/examples/degenerate-bevel-weld/degenerate_bevel_weld.py
@@ -150,10 +150,10 @@ def export_glb(ob, path):
     return path
 
 
-def check():
+def check(both_safe=False):
     tmp = tempfile.mkdtemp(prefix="bevelweld_")
     safe = beveled_box(DIMS, SAFE_OFFSET)
-    degen = beveled_box(DIMS, DEGEN_OFFSET)
+    degen = beveled_box(DIMS, SAFE_OFFSET if both_safe else DEGEN_OFFSET)
 
     # --- 1. threshold: zero-area faces flip on at offset == half min dim ---
     safe_za = zero_area_count(safe)
@@ -391,11 +391,13 @@ def main():
     p.add_argument("--output", default=None, help="optional: render a still PNG here")
     p.add_argument("--engine", default="eevee", choices=("eevee", "cycles"),
                    help="render engine for --output (cycles for GPU-less hosts)")
+    p.add_argument("--both-safe", action="store_true",
+                   help="bevel the degenerate box at the safe offset (must fail)")
     args = p.parse_args(argv)
 
     print(f"binary version: {bpy.app.version} ({bpy.app.version_string})")
     bpy.ops.wm.read_factory_settings(use_empty=True)
-    code = check()
+    code = check(both_safe=args.both_safe)
     if code:
         return code
 
diff --git a/examples/gn-instance-grid/README.md b/examples/gn-instance-grid/README.md
index 9038413..8e54115 100644
--- a/examples/gn-instance-grid/README.md
+++ b/examples/gn-instance-grid/README.md
@@ -19,11 +19,31 @@ counts fail.
 # Cheap correctness check (no render) — the CI check:
 blender --background --python gn_instance_grid.py --
 
+# Falsifier: 1×1 grid. Must exit non-zero (corner instance missing).
+blender --background --python gn_instance_grid.py -- --one-cell
+
 # Also render a still (EEVEE on a GPU host; use --engine cycles on GPU-less hosts):
 blender --background --python gn_instance_grid.py -- --output grid.png
 blender --background --python gn_instance_grid.py -- --output grid.png --engine cycles
 ```
 
-It exits non-zero on failure (wrong carrier, topology mismatch, missing material, or
-misplaced corner). The `blender-smoke` workflow runs the check on Blender 5.2 LTS and
-4.5 LTS.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | Carrier vertex count ≠ 1 |
+| 4 | Corner instance vert count ≠ 8 (`--one-cell` lands here) |
+| 5 | Evaluated topology ≠ 72 verts / 54 faces |
+| 6 | Set Material did not carry Lime |
+| 7 | Corner instance center off the closed-form grid point |
+| 8 | `--output` produced no file |
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--one-cell`.
diff --git a/examples/gn-instance-grid/gn_instance_grid.py b/examples/gn-instance-grid/gn_instance_grid.py
index 8897932..ba9ffa2 100644
--- a/examples/gn-instance-grid/gn_instance_grid.py
+++ b/examples/gn-instance-grid/gn_instance_grid.py
@@ -8,10 +8,15 @@
 left as empty instance geometry, and that a corner instance sits at its
 closed-form grid coordinate.
 
+``--one-cell`` builds a 1x1 grid and still asserts 3x3 realized topology,
+so the corner instance is missing. That is the falsifier
+(``--same-axis`` in export-preset-axis).
+
 By default it runs only the correctness check (no render) — the CI smoke
 check. Pass --output to also render a still:
 
     blender --background --python gn_instance_grid.py --                 # check only
+    blender --background --python gn_instance_grid.py -- --one-cell      # must fail
     blender --background --python gn_instance_grid.py -- --output g.png  # + render
 """
 import bpy, bmesh, sys, os, math, argparse
@@ -31,7 +36,7 @@
 CORNER_CENTER = (GRID_HALF, GRID_HALF, CUBE_SIZE / 2)
 
 
-def build_instance_grid_tree(material=None):
+def build_instance_grid_tree(material=None, grid_x=GRID_X, grid_y=GRID_Y):
     tree = bpy.data.node_groups.new("InstanceGrid", 'GeometryNodeTree')
     # generative: no Group Input — the tree owns the geometry
     tree.interface.new_socket(
@@ -42,8 +47,8 @@ def build_instance_grid_tree(material=None):
     grid = tree.nodes.new('GeometryNodeMeshGrid')
     grid.inputs["Size X"].default_value = GRID_SIZE
     grid.inputs["Size Y"].default_value = GRID_SIZE
-    grid.inputs["Vertices X"].default_value = GRID_X
-    grid.inputs["Vertices Y"].default_value = GRID_Y
+    grid.inputs["Vertices X"].default_value = grid_x
+    grid.inputs["Vertices Y"].default_value = grid_y
 
     cube = tree.nodes.new('GeometryNodeMeshCube')
     cube.inputs["Size"].default_value = (CUBE_SIZE, CUBE_SIZE, CUBE_SIZE)
@@ -75,7 +80,7 @@ def build_instance_grid_tree(material=None):
     return tree
 
 
-def build():
+def build(one_cell=False):
     bpy.ops.wm.read_factory_settings(use_empty=True)
     # carrier mesh is unused by the generative tree; one vertex is enough
     me = bpy.data.meshes.new("Carrier")
@@ -89,7 +94,8 @@ def build():
     bsdf.inputs["Base Color"].default_value = (0.22, 0.95, 0.06, 1.0)  # lime
     bsdf.inputs["Roughness"].default_value = 0.22
 
-    tree = build_instance_grid_tree(material=mat)
+    gx = gy = 1 if one_cell else GRID_X
+    tree = build_instance_grid_tree(material=mat, grid_x=gx, grid_y=gy)
     mod = obj.modifiers.new("instance_grid", 'NODES')
     mod.node_group = tree
     return obj, mat
@@ -246,9 +252,11 @@ def main():
     p.add_argument("--output", default=None, help="optional: render a still PNG here")
     p.add_argument("--engine", default="eevee", choices=("eevee", "cycles"),
                    help="render engine for --output (cycles for GPU-less hosts)")
+    p.add_argument("--one-cell", action="store_true",
+                   help="instance a 1x1 grid (must fail)")
     args = p.parse_args(argv)
 
-    obj, _mat = build()
+    obj, _mat = build(one_cell=args.one_cell)
     code = check(obj)
     if code:
         return code
diff --git a/examples/gn-sdf-remesh/README.md b/examples/gn-sdf-remesh/README.md
index df8e6da..98d4b99 100644
--- a/examples/gn-sdf-remesh/README.md
+++ b/examples/gn-sdf-remesh/README.md
@@ -22,12 +22,29 @@ output — this example does that and asserts the material survives onto the eva
 # Cheap correctness check only (no render) — the CI smoke check:
 blender --background --python gn_sdf_remesh.py --
 
+# Falsifier: no SDF modifier. Must exit non-zero (evaluated == base).
+blender --background --python gn_sdf_remesh.py -- --no-sdf
+
 # Also render the remeshed result (EEVEE on a GPU host; --engine cycles on GPU-less hosts):
 blender --background --python gn_sdf_remesh.py -- --output remesh.png
 blender --background --python gn_sdf_remesh.py -- --output remesh.png --engine cycles
 ```
 
-By default it runs only the **frame-independent correctness check**: the depsgraph-evaluated
-vertex count must be > 0 AND differ from the base mesh (the remesh produced geometry). It
-exits non-zero on failure — the same check the `blender-smoke` workflow runs on Blender 5.2 LTS and
-4.5 LTS.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | SDF remesh produced no/unchanged geometry (`--no-sdf` lands here) |
+| 4 | `--output` produced no file |
+| 5 | Wrong-era EEVEE engine id was accepted |
+| 6 | Input material dropped by remesh |
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--no-sdf`.
diff --git a/examples/gn-sdf-remesh/gn_sdf_remesh.py b/examples/gn-sdf-remesh/gn_sdf_remesh.py
index 6ab360a..d811da5 100644
--- a/examples/gn-sdf-remesh/gn_sdf_remesh.py
+++ b/examples/gn-sdf-remesh/gn_sdf_remesh.py
@@ -11,6 +11,7 @@
 both builds.
 
     blender --background --python gn_sdf_remesh.py --                  # correctness check only
+    blender --background --python gn_sdf_remesh.py -- --no-sdf         # must fail
     blender --background --python gn_sdf_remesh.py -- --output r.png   # also render the result
     blender --background --python gn_sdf_remesh.py -- --output r.png --engine cycles  # GPU-less
 """
@@ -108,6 +109,8 @@ def main():
     p = argparse.ArgumentParser()
     p.add_argument("--output", default=None, help="optional: render the remeshed result to this PNG")
     p.add_argument("--engine", choices=["auto", "cycles"], default="auto")
+    p.add_argument("--no-sdf", action="store_true",
+                   help="skip attaching the SDF remesh modifier (must fail)")
     args = p.parse_args(argv)
 
     # EEVEE-id inversion witnessed for real: the OTHER era's id must be
@@ -125,7 +128,8 @@ def main():
     base = len(obj.data.vertices)
     src_mat = obj.data.materials[0] if obj.data.materials else None
     tree, link_valid = build_remesh_via_sdf(material=src_mat)
-    obj.modifiers.new("sdf", 'NODES').node_group = tree
+    if not args.no_sdf:
+        obj.modifiers.new("sdf", 'NODES').node_group = tree
     dg = bpy.context.evaluated_depsgraph_get(); ev = obj.evaluated_get(dg)
     m = ev.to_mesh(); evc = len(m.vertices)
     mat_names = [mm.name for mm in m.materials if mm is not None]
diff --git a/examples/mesh-hygiene-audit/README.md b/examples/mesh-hygiene-audit/README.md
index 98670f5..2ab4ea2 100644
--- a/examples/mesh-hygiene-audit/README.md
+++ b/examples/mesh-hygiene-audit/README.md
@@ -49,11 +49,30 @@ thumbnail scale without faking annotation.
 
 ```bash
 blender --background --python mesh_hygiene_audit.py --
+blender --background --python mesh_hygiene_audit.py -- --inject-ngon
 blender --background --python mesh_hygiene_audit.py -- --output hygiene.png
 blender --background --python mesh_hygiene_audit.py -- --output hygiene.png --engine cycles
 ```
 
-Exits non-zero on failure. The `blender-smoke` workflow runs the check on
-Blender 5.2 LTS and 4.5 LTS. The `--output` render path additionally measures
-framing against the Layer 1 band via `examples/gallery_framing.py` (exit 10
-on violation) before writing the still.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it. `10` is the shared framing helper.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | Ngon present (`--inject-ngon` lands here) |
+| 4 | Loose vertices |
+| 5 | Non-manifold or boundary edges |
+| 6 | Zero-area faces |
+| 7 | Signed volume ≤ 0 |
+| 8 | Euler characteristic ≠ 2 |
+| 9 | `--output` produced no file |
+| 10 | Gallery framing violation |
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--inject-ngon`.
diff --git a/examples/mesh-hygiene-audit/mesh_hygiene_audit.py b/examples/mesh-hygiene-audit/mesh_hygiene_audit.py
index a31268f..ba9d6e4 100644
--- a/examples/mesh-hygiene-audit/mesh_hygiene_audit.py
+++ b/examples/mesh-hygiene-audit/mesh_hygiene_audit.py
@@ -15,6 +15,7 @@
 to also render a still:
 
     blender --background --python mesh_hygiene_audit.py --
+    blender --background --python mesh_hygiene_audit.py -- --inject-ngon
     blender --background --python mesh_hygiene_audit.py -- --output h.png
 """
 import bpy, bmesh, sys, os, math, argparse
@@ -577,10 +578,14 @@ def main():
         "--engine", default="eevee", choices=("eevee", "cycles"),
         help="render engine for --output",
     )
+    p.add_argument("--inject-ngon", action="store_true",
+                   help="dissolve one edge into an ngon (must fail)")
     args = p.parse_args(argv)
 
     print(f"binary version: {bpy.app.version} ({bpy.app.version_string})")
     sc, ob = build_scene()
+    if args.inject_ngon:
+        inject_defect(ob.data, "ngon")
     code = check(ob.data)
     if code:
         return code
diff --git a/examples/swatch-grid/README.md b/examples/swatch-grid/README.md
index 28778c0..c8cba9f 100644
--- a/examples/swatch-grid/README.md
+++ b/examples/swatch-grid/README.md
@@ -23,6 +23,9 @@ that mapping fails the example, not just the docs.
 # Cheap correctness check (materials + engine-id witness, no render):
 blender --background --python swatch_grid.py --
 
+# Falsifier: same RGB on every swatch. Must exit non-zero.
+blender --background --python swatch_grid.py -- --same-base
+
 # Render and pixel-verify with the build's EEVEE engine (needs a GPU/display):
 blender --background --python swatch_grid.py -- --output swatch.png
 
@@ -31,10 +34,27 @@ blender --background --python swatch_grid.py -- --output swatch.png
 blender --background --python swatch_grid.py -- --output swatch.png --engine cycles --samples 16 --width 960
 ```
 
-The script is deterministic and dependency-light (fixed camera and layout, no HDRI, no
-network). It **exits non-zero** on any failure, including a render that comes out uniformly
-black or without the expected six distinct swatch regions — the same honest check the CI
-smoke gate runs on Blender 5.2 LTS and 4.5 LTS.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it. `10` is the shared framing helper.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | Distinct swatch colors ≠ 6 (`--same-base` lands here); also render not six distinct regions |
+| 4 | `--output` produced no file |
+| 5 | Wrong-era EEVEE engine id was accepted |
+| 10 | Gallery framing violation |
+
+`--no-verify` was a skip-flag and has been removed. Pixel verification always
+runs when `--output` is passed.
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--same-base`.
 
 ## Verified
 
diff --git a/examples/swatch-grid/swatch_grid.py b/examples/swatch-grid/swatch_grid.py
index a1997d5..80e65d6 100644
--- a/examples/swatch-grid/swatch_grid.py
+++ b/examples/swatch-grid/swatch_grid.py
@@ -8,9 +8,12 @@
 on 4.2-4.5, and the chosen id is asserted against the build before rendering.
 
 By default it runs only the correctness check (no render) — the CI smoke check.
-Pass --output to also render and pixel-verify a still:
+Pass --output to also render and pixel-verify a still. ``--same-base`` writes
+the same RGB to every swatch and still asserts six distinct colors, so the
+count fails. That is the falsifier (``--same-axis`` in export-preset-axis).
 
     blender --background --python swatch_grid.py --                       # check only
+    blender --background --python swatch_grid.py -- --same-base            # must fail
     blender --background --python swatch_grid.py -- --output swatch.png
     blender --background --python swatch_grid.py -- --output s.png --engine cycles --samples 8 --width 640
 
@@ -183,6 +186,39 @@ def build_scene(mats):
     bpy.context.scene.world = world
 
 
+def swatch_rgb(mat):
+    for node in mat.node_tree.nodes:
+        if node.type == "BSDF_PRINCIPLED":
+            c = node.inputs["Base Color"].default_value
+            return (round(c[0], 4), round(c[1], 4), round(c[2], 4))
+        if node.type == "EMISSION":
+            c = node.inputs["Color"].default_value
+            return (round(c[0], 4), round(c[1], 4), round(c[2], 4))
+    return None
+
+
+def flatten_swatch_colors(mats):
+    gray = (0.5, 0.5, 0.5, 1.0)
+    for mat in mats:
+        for node in mat.node_tree.nodes:
+            if node.type == "BSDF_PRINCIPLED":
+                node.inputs["Base Color"].default_value = gray
+            elif node.type == "EMISSION":
+                node.inputs["Color"].default_value = gray
+
+
+def check_distinct_swatches(mats):
+    colors = [swatch_rgb(m) for m in mats]
+    if len(set(colors)) != MATERIAL_COUNT:
+        print(
+            f"ERROR: distinct swatch colors {len(set(colors))} != "
+            f"{MATERIAL_COUNT} (got {colors})",
+            file=sys.stderr,
+        )
+        return 3
+    return 0
+
+
 def verify_png(path):
     """Honest capture: not uniformly black AND distinct swatch regions == MATERIAL_COUNT."""
     img = bpy.data.images.load(path)
@@ -210,12 +246,18 @@ def main():
                    help="auto/eevee use the version-correct EEVEE id; cycles for GPU-less hosts")
     p.add_argument("--samples", type=int, default=32)
     p.add_argument("--width", type=int, default=1280)
-    p.add_argument("--no-verify", action="store_true")
+    p.add_argument("--same-base", action="store_true",
+                   help="write the same RGB to every swatch (must fail)")
     args = p.parse_args(argv)
 
     # Empty the factory file FIRST so the materials we create below survive.
     bpy.ops.wm.read_factory_settings(use_empty=True)
     mats, specular_socket = build_materials()
+    if args.same_base:
+        flatten_swatch_colors(mats)
+    dcode = check_distinct_swatches(mats)
+    if dcode:
+        return dcode
     build_scene(mats)
 
     sc = bpy.context.scene
@@ -271,15 +313,14 @@ def main():
         return 4
     print(f"rendered {args.output} with {render_engine} ({os.path.getsize(args.output)} bytes)")
 
-    if not args.no_verify:
-        gmax, regions = verify_png(args.output)
-        non_black = gmax > 0.05
-        regions_ok = regions == MATERIAL_COUNT
-        print(f"verify: max_pixel={gmax:.3f} non_black={non_black} "
-              f"distinct_regions={regions} materials={MATERIAL_COUNT} ok={regions_ok}")
-        if not (non_black and regions_ok):
-            print("ERROR: render failed verification (black or wrong region count)", file=sys.stderr)
-            return 3
+    gmax, regions = verify_png(args.output)
+    non_black = gmax > 0.05
+    regions_ok = regions == MATERIAL_COUNT
+    print(f"verify: max_pixel={gmax:.3f} non_black={non_black} "
+          f"distinct_regions={regions} materials={MATERIAL_COUNT} ok={regions_ok}")
+    if not (non_black and regions_ok):
+        print("ERROR: render failed verification (black or wrong region count)", file=sys.stderr)
+        return 3
     print("swatch-grid OK")
     return 0
 
diff --git a/examples/triangulate-tangents/README.md b/examples/triangulate-tangents/README.md
index d03f748..4ef654f 100644
--- a/examples/triangulate-tangents/README.md
+++ b/examples/triangulate-tangents/README.md
@@ -59,14 +59,32 @@ made visible.
 # Cheap correctness check (no render) — the CI check:
 blender --background --python triangulate_tangents.py --
 
+# Falsifier: every UV at (0, 0). Must exit non-zero (authored UV closed form).
+blender --background --python triangulate_tangents.py -- --zero-uv
+
 # Also render a still (EEVEE on a GPU host; use --engine cycles on GPU-less hosts):
 blender --background --python triangulate_tangents.py -- --output buckler.png
 blender --background --python triangulate_tangents.py -- --output buckler.png --engine cycles
 ```
 
-It exits non-zero on failure (topology drift, reallocated UV layer,
-non-orthonormal basis, bitangent-convention drift, formula excursion, or a
-flip inside a smooth field). The `blender-smoke` workflow runs the check on
-Blender 5.2 LTS and 4.5 LTS.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it. `10` is the shared framing helper.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | Loop-triangle count ≠ closed form |
+| 4 | Re-fetched UV layer drifted from the polar field (`--zero-uv` lands here) |
+| 5 | Tangent basis not orthonormal |
+| 6 | Bitangent ≠ sign × (n × t) |
+| 7 | Tangents deviate from the edge/UV closed form |
+| 8 | Flipped tangent inside a clean triangle |
+| 10 | Gallery framing violation; also `--output` produced no file |
 
-The `--output` render path additionally measures framing against the Layer 1 band via `examples/gallery_framing.py` (exit 10 on violation) before writing the still.
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--zero-uv`.
diff --git a/examples/triangulate-tangents/triangulate_tangents.py b/examples/triangulate-tangents/triangulate_tangents.py
index c91af01..6072800 100644
--- a/examples/triangulate-tangents/triangulate_tangents.py
+++ b/examples/triangulate-tangents/triangulate_tangents.py
@@ -31,6 +31,7 @@
 check. Pass --output to also render a still:
 
     blender --background --python triangulate_tangents.py --                 # check only
+    blender --background --python triangulate_tangents.py -- --zero-uv       # must fail
     blender --background --python triangulate_tangents.py -- --output s.png  # + render
 """
 import bpy, bmesh, sys, os, math, argparse
@@ -428,10 +429,16 @@ def main():
     p.add_argument("--output", default=None, help="optional: render a still PNG here")
     p.add_argument("--engine", default="eevee", choices=("eevee", "cycles"),
                    help="render engine for --output (cycles for GPU-less hosts)")
+    p.add_argument("--zero-uv", action="store_true",
+                   help="write every UV to (0, 0) (must fail)")
     args = p.parse_args(argv)
 
     bpy.ops.wm.read_factory_settings(use_empty=True)
     obj, n_dome = build_buckler()
+    if args.zero_uv:
+        uv = obj.data.uv_layers["UVMap"]
+        for loop in uv.data:
+            loop.uv = (0.0, 0.0)
     code = check(obj, n_dome)
     if code:
         return code
diff --git a/examples/uv-layer-grid/README.md b/examples/uv-layer-grid/README.md
index b9b5741..1200314 100644
--- a/examples/uv-layer-grid/README.md
+++ b/examples/uv-layer-grid/README.md
@@ -62,11 +62,38 @@ the pixels prove the flat-vs-checker split (measured on 5.1.2: hazard spread
 # Cheap correctness check (no render) — the CI check:
 blender --background --python uv_layer_grid.py --
 
+# Falsifier: pre-create a UV layer on the silent-no-op probe. Must exit non-zero.
+blender --background --python uv_layer_grid.py -- --precreate-on-hazard
+
 # Also render a still (EEVEE on a GPU host; use --engine cycles on GPU-less hosts):
 blender --background --python uv_layer_grid.py -- --output uv.png
 blender --background --python uv_layer_grid.py -- --output uv.png --engine cycles
 ```
 
-It exits non-zero on failure and prints every measured error and tolerance on
-success, so CI logs carry the numbers. The `blender-smoke` workflow runs the
-check on Blender 5.2 LTS and 4.5 LTS.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it. This example does not call the gallery framing helper; `10`
+is the missing-render-file check.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 3 | Silent-no-op hazard gone (`--precreate-on-hazard` lands here) |
+| 4 | Grid topology drifted |
+| 5 | Pre-create + calc_uvs did not persist one UV layer |
+| 6 | calc_uvs closed-form error |
+| 7 | Mesh UV round-trip error |
+| 8 | calc_uvs=False unexpectedly created a UV layer |
+| 9 | Explicit UV assignment error |
+| 10 | `--output` produced no file |
+| 11 | Broken panel is not flat |
+| 12 | Broken panel is not texel-(0,0) teal |
+| 13 | Repaired panel is not a checker |
+| 14 | Broken and repaired panels render identically |
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--precreate-on-hazard`.
diff --git a/examples/uv-layer-grid/uv_layer_grid.py b/examples/uv-layer-grid/uv_layer_grid.py
index 908b045..25ae49f 100644
--- a/examples/uv-layer-grid/uv_layer_grid.py
+++ b/examples/uv-layer-grid/uv_layer_grid.py
@@ -8,14 +8,14 @@
 
 The check proves the silent no-op, then the pre-create + `calc_uvs=True`
 repair path against a closed-form UV grid, and an explicit loop-assignment
-fallback that does not depend on `calc_uvs` at all. Pass --output to also
-render a still that stages the broken (flat) panel beside the repaired
-(checker) panel — and then *witnesses the render itself*: the saved PNG is
-read back and probed at each panel's projected center, asserting the broken
-panel is one flat teal (texel (0,0)) while the repaired panel carries both
-checker colors. If the UV contract failed, the pixels would say so:
+fallback that does not depend on `calc_uvs` at all.
+
+``--precreate-on-hazard`` creates the UV layer on the silent-no-op probe
+and still asserts zero layers, so the hazard check fails. That is the
+falsifier (``--same-axis`` in export-preset-axis).
 
     blender --background --python uv_layer_grid.py --
+    blender --background --python uv_layer_grid.py -- --precreate-on-hazard
     blender --background --python uv_layer_grid.py -- --output uv.png
 """
 import bpy, bmesh, sys, os, math, argparse
@@ -46,13 +46,15 @@ def max_uv_err(bm, uv_layer):
     return err
 
 
-def check():
+def check(precreate_on_hazard=False):
     bpy.ops.wm.read_factory_settings(use_empty=True)
 
     # --- 1. The hazard: calc_uvs=True is a silent no-op without a UV layer ---
     me_bad = bpy.data.meshes.new("NoPreUV")
     bm = bmesh.new()
     try:
+        if precreate_on_hazard:
+            bm.loops.layers.uv.new("UVMap")
         bmesh.ops.create_grid(
             bm, x_segments=SEG, y_segments=SEG, size=SIZE, calc_uvs=True,
         )
@@ -506,9 +508,11 @@ def main():
         "--engine", default="eevee", choices=("eevee", "cycles"),
         help="render engine for --output (cycles for GPU-less hosts)",
     )
+    p.add_argument("--precreate-on-hazard", action="store_true",
+                   help="pre-create a UV layer on the silent-no-op probe (must fail)")
     args = p.parse_args(argv)
 
-    code = check()
+    code = check(precreate_on_hazard=args.precreate_on_hazard)
     if code != 0:
         return code
     if args.output:
diff --git a/examples/wave-displace/README.md b/examples/wave-displace/README.md
index aab4b3c..d984e02 100644
--- a/examples/wave-displace/README.md
+++ b/examples/wave-displace/README.md
@@ -25,10 +25,28 @@ to `examples/gallery_framing.py`, call it with
 # Cheap correctness check (no render) — the CI check:
 blender --background --python wave_displace.py --
 
+# Falsifier: skip the foreach_set displacement. Must exit non-zero (z-span).
+blender --background --python wave_displace.py -- --flat
+
 # Also render a still (EEVEE on a GPU host; use --engine cycles on GPU-less hosts):
 blender --background --python wave_displace.py -- --output wave.png
 blender --background --python wave_displace.py -- --output wave.png --engine cycles
 ```
 
-It exits non-zero on failure (span wrong, or any vertex off the closed form). The
-`blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS.
+## Exit codes
+
+Per-script sequential checks. `9` is a valid check code; there is no rule
+against it.
+
+| Code | Meaning |
+| --- | --- |
+| 0 | Success |
+| 1 | Uncaught exception (FATAL wrapper) |
+| 2 | argparse / usage |
+| 4 | Z-span not in the closed-form band (`--flat` lands here) |
+| 5 | A vertex is off the closed-form wave |
+| 6 | `--output` produced no file |
+
+The `blender-smoke` workflow runs the check on Blender 5.2 LTS and 4.5 LTS
+(5.1 on the weekly cron, the `needs-5.1` PR label, or manual dispatch).
+Smoke does not pass `--output` or `--flat`.
diff --git a/examples/wave-displace/wave_displace.py b/examples/wave-displace/wave_displace.py
index 677918f..8602120 100644
--- a/examples/wave-displace/wave_displace.py
+++ b/examples/wave-displace/wave_displace.py
@@ -12,6 +12,7 @@
 check. Pass --output to also render a still:
 
     blender --background --python wave_displace.py --                 # check only
+    blender --background --python wave_displace.py -- --flat          # must fail
     blender --background --python wave_displace.py -- --output w.png  # + render
 """
 import bpy, bmesh, sys, os, math, argparse
@@ -133,10 +134,12 @@ def main():
     p.add_argument("--output", default=None, help="optional: render a still PNG here")
     p.add_argument("--engine", default="eevee", choices=("eevee", "cycles"),
                    help="render engine for --output (cycles for GPU-less hosts)")
+    p.add_argument("--flat", action="store_true",
+                   help="skip the foreach_set displacement (must fail)")
     args = p.parse_args(argv)
 
     obj = build_grid()
-    n = displace(obj.data)
+    n = len(obj.data.vertices) if args.flat else displace(obj.data)
     code = check(obj, n)
     if code:
         return code