@deepblend/dsh-blender-preset 0.8.1 → 0.10.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@deepblend/dsh-blender-preset",
-  "version": "0.8.1",
+  "version": "0.10.0",
   "type": "module",
   "description": "DeepBlend Agent-preset plane: ships the DeepBlend Studio agent presets and the deployer row that installs them into the DSH home.",
   "main": "lib/index.js",

package/presets/deepblend/skills/deepblend-studio/SKILL.md

@@ -65,9 +65,78 @@ twice the body radius; the whole object's bounding box includes the handle.
 Declared wall thickness and sampled technical checks do not certify a global
 minimum wall or manufacturing tolerance.
 
+For a dimension or assembly check, read a few relevant saved parts with
+`blender_scene_get {projectId, revision, geometryEntityIds, geometryFrame}`.
+Use IDs from the returned scene and the explicit frame of the preview being checked.
+This measures the actual saved native mesh without a render or another revision;
+prefer that small receipt to repeated full-scene searches when you need bounds or
+transforms. Keep its source revision and checkpoint hash with the result. World-axis
+spans include rotation, handles and protrusions: they do not prove body diameter,
+wall thickness, clearance or quality. If a part is unavailable or the runtime refuses
+inspection, state that limit instead of substituting generator settings as measurements.
+After changing dimensions, check the affected parts and important attachments again,
+then use actual preview pixels to judge the finish and joins.
+
+Do not turn the difference between shell and liner outer bounds into an assembly
+gap. It includes shell material and depends on curvature and orientation; it is
+not the distance from the shell's inner surface to the liner. An unchanged height,
+position or scale also cannot establish preserved wall thickness or clearance.
+When those conditions are requested, distinguish the recipe's linked design rule
+from actual surface measurements. Without a separate supported surface check,
+report wall thickness and clearance as unmeasured; do not give a numeric gap or
+claim "basically preserved" from bounding dimensions or preview pixels.
+
+For actual local surface evidence, the same tool optionally accepts
+`geometrySurfaceSamples` (1–8). Each sample supplies `sampleId`, a selected
+`entityId`, an outside `originMeters`, a non-zero `direction`, and explicit
+`searchDistanceMm` (0.01–500); optional `targetEntityId` must be another selected
+part. Choose seeds from actual saved bounds and the requested region, not from
+declared thickness or gap values. The runtime follows the first entry face's
+normal to its first exit and, if requested, the target's first entry. Keep seeds,
+frame, part IDs and source SHA with the receipt; use the same explicit seeds to
+compare saved revisions. A failed sample is unavailable, never zero.
+
+Use `geometryOnly:true` with the native selection for a compact measurement receipt;
+it keeps the complete verified objects, samples and source identity while omitting
+the scene catalogue. Read the scene separately when selecting parts or seeds.
+It cannot be combined with `full:true`.
+
+For before/after comparisons, supply `geometryCompareRevision` with the old saved
+revision and the same explicit samples on the current read. The tool verifies both
+sources and computes `geometryComparison.samples` changes (current minus reference)
+and maxima over **all** comparable requested samples. Chord-only samples still count
+in the chord maximum. Unavailable pairs are omitted from numeric changes, never zero.
+Use these computed summaries rather than mental arithmetic. Preserve sample IDs,
+coverage counts and unavailable rows; round only for display. Do not assert left/right
+symmetry or an explanation for a change from this summary. A world seed does not
+necessarily hit the same material point across revisions. If the comparison is
+unavailable, report the raw evidence and limit the claim instead of guessing maxima.
+The comparison also computes percent changes against each reference value. Quote
+the separate chord and separation percentage maxima exactly (round only for display).
+Do not apply a chord percentage to the gap, derive percentages from rounded table
+values, or introduce a combined relative bound. If no percentage receipt exists,
+omit the percentage claim. Do not equate a measured change with user approval of
+its size; state the observed values and coverage, and leave unmeasured conditions open.
+
+`normalSurfaceChordMm` runs from source entry to source exit. Target
+`normalRaySeparationMm` runs from source **EXIT** to target entry along that
+inward normal, excluding the source chord. Describe those two surfaces explicitly;
+do not label it as source outer surface to target entry or as global free clearance.
+Triangle-face normals differ from shading normals. These samples do not prove
+global minimum wall thickness, whole-assembly free clearance, no collision,
+manufacturing safety or visual quality. Other scene objects and mesh
+self-intersections are not certified. Chord/target searches each stop at 100 mm;
+unsupported topology, overlap, occlusion and missed rays must remain explicit.
+If no supported samples cover a requested condition, keep it unmeasured.
+
 Inspect the saved preview needed for the brief; add a targeted view if it does
 not show the requested features. Once the requested native file and preview
 are saved, conclude with the project, revision and actual delivery paths.
+Use the current operation's returned project-relative paths and identify the project
+and revision. A historical producer path does not establish the current store root.
+Do not guess absolute paths from old metadata. If a generic presentation tool cannot
+register a project-relative file, report its path and use the saved workbench preview;
+do not invent a successful attachment or search unrelated stores to fabricate one.
 Separate verified dimensions/topology from appearance observations. An
 unavailable vision reviewer is a limitation to report, not a reason to repeat
 unchanged renders or to claim artistic approval. Additional artistic review
@@ -297,6 +366,10 @@ hard to get wrong:
 - **The score and the issues are measured, not reported.** The host computes them
   deterministically from the rendered pixels — composition, exposure, occlusion. The
   same scene scores the same twice, and a contract test asserts that.
+  Keep measured open issues visible in the handover. An unexpected threshold or
+  plausible explanation does not establish a false positive. Distinguish a possible
+  limitation from a verified cause; dismiss an issue only with separate evidence
+  from the relevant view and object.
 - **You are the only one who can see meaning.** The measurements cannot tell a watch
   dial that is *supposed* to be bright from one that is blown out, or a screen that is
   *supposed* to be hidden from one that is occluded by mistake. When a scene measures