Slice 13 — mRBFShape + mRBFDrawOverride (Path B) by 891458249 · Pull Request #11 · 891458249/RBF_MAX

891458249 · 2026-04-21T21:29:42Z

Summary

Phase 2B opening slice: Viewport 2.0 visualization for trained mRBFNode centers.

Architecture (Path B) — mRBFNode stays a pure kDependNode; visualization is carried by a new auxiliary locator mRBFShape connected via message attribute. mRBFDrawOverride is registered against the shape's classification.

mRBFNode (kDependNode, Phase 2A unchanged)
    │  .message ─► mRBFShape.sourceNode
    ▼
mRBFShape (kLocatorNode, classification drawdb/geometry/rbfmax/mRBFShape)
    │
    ▼
mRBFDrawOverride → MUIDrawManager::sphere() per center

Path A → B retrospective

Original spec attempted reparenting mRBFNode itself to MPxLocatorNode with 6-arg registerNode(…, kLocatorNode, &classification). Maya returned kFailure: "Unexpected Internal Failure" on both 2022 and 2025. 4+ hours of systematic assumption elimination (Rule 4) ruled out: classification format (3-level/2-level), MString storage (stack/static/anon-namespace), devkit provenance, registration order, Maya runtime version, output-only kString attr subset, C++ inheritance (via static_assert), and stale build artifacts. An out-of-repo reducer (MinLocator3.cpp) empirically pinned the failure to MPxLocatorNode + MFnTypedAttribute(MFnData::kString). Root Maya-internal cause unresolved; strategic retreat to Path B — a separate locator whose attribute set is strictly numeric + message. Full retrospective in DEVLOG.md Slice 13 entry; proposed Rule 5 candidate for Phase 2B memory: prefer canonical Autodesk sample patterns over ambitious deviations.

Deliverables

Phase 1 additive: RBFInterpolator::centers() const getter + 1 D-group test
IDrawSink / DrawCall / emit_centers_draw_calls adapter + 8 E-group tests (Maya-free)
mRBFNode::is_loaded() + centers_for_viewport() (read-only, no DG touch)
mRBFShape : MPxLocatorNode, typeId 0x00013A01 — attrs: sourceNode (message) + drawEnabled (bool) + sphereRadius (double)
mRBFDrawOverride : MPxDrawOverride — 2-hop resolve shape → mRBFNode, MUIDrawManager::sphere per center
plugin_main.cpp — registers all of the above with atomic rollback
tests/smoke/smoke_viewport.py — 7-step mayapy contract
DEVLOG.md + maya_node/README.md — Path A→B retrospective + Viewport 2.0 usage

Validation status

Check	Result
Adapter + Phase 1 tests (`build-adapter`, ctest)	163/163 passed
Maya 2022 plugin build (`C:/SDK/Maya2022/devkitBase`)	0 warn / 0 err
Maya 2025 plugin build (`C:/SDK/Maya2025/devkitBase`)	0 warn / 0 err
Maya 2022 × 4 smokes (hello/predict/train/viewport)	all PASS
Maya 2025 × 4 smokes (hello/predict/train/viewport)	all PASS
Phase 1 pure regression (`build`, ctest)	138/138 passed

Invariant changes

Cross-version .mll bit-identity is NO LONGER expected — Slice 13 adds OpenMayaRender + OpenMayaUI link libs which diverge across 2022/2025 devkits. The Phase 2A "bit-identical across versions" lemma applied only when the ABI surface was OpenMaya only; explicitly retracted in DEVLOG.md.
Phase 2A contract for mRBFNode fully preserved — no change to its compute, attributes, type (still kDependNode), or initialize().

Tech-debt added

R-44: MPxLocatorNode + MFnTypedAttribute(kString) empirical failure mode on Maya 2022 and 2025; root cause unresolved. Mitigation: Slice 14/15 must keep mRBFShape attrs numeric + message only.
T-16: Slice 13 requires user to connectAttr mRBFNode.message mRBFShape.sourceNode manually. Consider a convenience rbfmaxAttachShape command in a later slice.
T-17: centers_for_viewport() returns positions only; Slice 14 (heatmap) will need per-center colors.
T-18: Fixed (-10, +10) bbox on draw override; tighten in Slice 14+.

Test plan

Phase 2B Slice 13 prerequisite: Viewport 2.0 DrawOverride needs read-only access to the trained centers matrix for visualisation, mirrors the Slice 11 pattern where kernel_params() getter closed the aKernelType attribute gap without forcing the node to re-parse saved JSON. New public API (additive, noexcept, Maya-free, engine-agnostic): const MatrixX& centers() const noexcept; Contract: returns fit_result_.centers (owned by the interpolator since fit() or load() populated it); before the first successful fit/load returns a default-constructed 0x0 matrix. Gate on is_fitted() for defined row/col counts. No behavioural change to fit / predict / save / load / clone / ScratchPool paths. Test: new TEST(RBFInterpolatorState, CentersGetterReflectsFit) in category D (state queries; count goes 3 → 4). Reuses the hand- crafted 4-corner unit-square fixture shared with KernelParamsReflectsFit — trivially solvable at λ=1e-6, Gaussian, no polynomial tail. Asserts row/col count and element-wise EXPECT_DOUBLE_EQ equality against the input matrix. Phase 1 regression now reports **138/138 green** (was 137/137). Second Phase 1 additive amendment in Phase 2 under the Slice 11 precedent (additive const getter + test). Section G allow-list from Slice 11 covers this case verbatim; no new DEVLOG amendment needed for this specific commit. Local verification (MSVC 19.44 Release): 138/138 green, 12.97 s. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Phase 2B Slice 13 piece 2 of 5. Adds a Maya-free abstraction for accumulating Viewport 2.0 draw primitives + one orchestrator that emits one sphere per trained center. Deliberately does NOT wire the DrawOverride through this sink in Slice 13. The current draw topology (single filled-sphere pass per mRBFNode instance) does not justify the indirection; calling MUIDrawManager directly from addUIDrawables keeps the native side thin. The abstraction lands now because Slice 14 (heatmap) and Slice 15 (X-ray) will need to share orchestration — better to ship the shape of the contract early and grow tests around it than to retrofit on emergency. Files: * maya_node/include/rbfmax/maya/draw_sink.hpp - DrawCall POD (type enum + p0/p1/radius/color, default- constructible with sane zeros) - IDrawSink abstract interface (begin / emit / end, all noexcept per contract) - emit_centers_draw_calls free function signature * maya_node/src/draw_sink_core.cpp - emit_centers_draw_calls body. Always calls begin/end exactly once even for empty input so downstream sinks get well-defined transaction boundaries. Tests (E group, 8 new TESTs, adapter suite): E1 MockSinkRecordsEmittedCalls — sink contract smoke test E2 EmitCentersSimpleCount — 4 in, 4 kSphere out E3 EmitCentersColorPropagates — color argument preserved E4 EmitCentersRadiusPropagates — radius argument preserved E5 EmitCentersEmptyInput — begin/end still balanced E6 EmitCentersCoordsPreserved — element-wise bit-exact E7 EmitCentersBeginEndBalanced — 10 centers, still one begin/end pair E8 DrawCallDefaultsAreSane — default ctor gives kPoint at origin, radius 0, white; anchors kSeedS13 = 0xF5BFACu (reserved for future randomised E-group tests) Both draw_sink_core.cpp and the E tests link into the adapter GTest target, so the whole accumulator is covered without Maya being available — matches the Slice 10A/11/12 discipline of keeping orchestration logic Maya-free. Local verification (MSVC 19.44 Release): Step 1 adapter + Phase 1 combined: 163/163 green, 13.93 s (138 Phase 1 + 3 H + 6 C + 8 D + 8 E) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…or_viewport) Phase 2B Slice 13 Path B — groundwork for the mRBFDrawOverride that will be registered on the auxiliary mRBFShape locator (next commit). Both accessors are read-only and touch only interp_; Phase 2A compute / try_load / attribute set are unchanged. * is_loaded() -> bool, noexcept: interp_ != nullptr * centers_for_viewport() -> std::vector<MPoint>: first-3-dim projection of RBFInterpolator::centers() with zero-pad D<3 and truncation D>3 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Phase 2B Slice 13 Path B — new auxiliary locator that hosts the Viewport 2.0 draw override (next commit). mRBFNode stays a pure kDependNode; mRBFShape connects via message attribute: mRBFNode.message -> mRBFShape.sourceNode Attributes (deliberately numeric + message only — no MFnTypedAttribute of MFnData::kString, see DEVLOG Slice 13 Path A retrospective): * sourceNode — message, writable * drawEnabled — bool, default true * sphereRadius — double, default 0.05, min 0.001, softMax 1.0 typeId 0x00013A01 (monotonic after mRBFNode's 0x00013A00). CMake: adds OpenMayaUI to find_package COMPONENTS and mrbf_shape.cpp to the plugin target. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Phase 2B Slice 13 Path B — Viewport 2.0 integration. mRBFDrawOverride is registered against mRBFShape's classification "drawdb/geometry/rbfmax/mRBFShape". prepareForDraw upstreams via shape.sourceNode's message connection to resolve the connected mRBFNode, then reads is_loaded() + centers_for_viewport(). Any failure along the chain yields empty draw data (no crash, no warning). addUIDrawables emits one MUIDrawManager::sphere per center inside a beginDrawable/endDrawable pair. plugin_main.cpp gains a 6-arg registerNode for mRBFShape with MPxNode::kLocatorNode + classification pointer, and a MDrawRegistry::registerDrawOverrideCreator call keyed on that same classification. mRBFNode's 4-arg registerNode (Slice 10A kDependNode form) is unchanged. Mirror unwind + atomic rollback on any failure. smoke_viewport.py — 7-step Path B contract: 0. loadPlugin OK 1. mRBFNode in allNodeTypes(), no drawdb classification 2. mRBFShape in allNodeTypes() 3. mRBFShape classification matches DrawOverride literal 4. mRBFNode.message -> mRBFShape.sourceNode connect OK 5. Phase 2A predict round-trip on mRBFNode 6. isLoaded + nCenters state attrs reflect loaded interp 7. mRBFShape drawEnabled + sphereRadius round-trip Validated on Maya 2022 (devkit C:/SDK/Maya2022/devkitBase) and Maya 2025 (devkit C:/SDK/Maya2025/devkitBase); both 4-smoke suites exit 0 in each version. Adapter + Phase 1 tests 163/163. Phase 1 pure regression 138/138. CMake: adds OpenMayaRender to find_package COMPONENTS and mrbf_draw_override.cpp to the plugin target. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…0 skeleton DEVLOG.md: adds a ~180-line Slice 13 entry covering scope, deliverables, design decisions, the full 4+ hour Path A failure analysis (assumption-elimination trajectory: classification format, MString storage, devkit, registration order, Maya version, output- string attrs, inheritance static_assert, stale artifacts), the Path B architecture diagram, a Rule 5 candidate for Phase 2B memory (prefer canonical Autodesk sample pattern over ambitious deviations), R-09 typeId self-checks, validation outcomes, and a tech-debt register (R-44, T-16, T-17, T-18). maya_node/README.md: adds a "Viewport 2.0 visualization" section with the mRBFShape attribute table, the create-and-connect workflow, dimension projection / unloaded-state / multiple-shape notes; updates the Files table with the 4 new headers, 4 new implementations, the 2 new tests, and smoke_viewport.py. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Companion to Slice 13 — lets end users install/uninstall the rbfmax plugin by dragging a Python file into the Maya viewport. Files added ----------- - installer/drag_drop_install.py — onMayaDroppedPythonFile hook + install / uninstall via Maya Module System (.mod descriptor written into the user's ~/Documents/maya/modules). Single- version install covers Maya 2022 and 2025 via per-version MAYAVERSION blocks in the .mod file. - installer/package.py — copies built .mll from build-maya-<ver>/bin/Release/ into installer/plug-ins/<ver>/. Strict fail-fast build-missing preflight: exits 1 with a full cmake recipe if either version's .mll is absent. - installer/README.md — Chinese user guide covering both the end- user and developer flows, plus a troubleshooting section. - docs/install.md — repository-level user entry pointing at the installer (so GitHub viewers see the Maya install path without having to download the zip first). Files updated ------------- - README.md — adds a blockquote under Quick Start pointing Maya end users at docs/install.md. - .gitignore — excludes installer/plug-ins/2022/*.mll and installer/plug-ins/2025/*.mll (large build artifacts; users populate them via package.py). Validation ---------- mayapy install + uninstall smoke cycle on Maya 2022 and 2025: - package.py copies .mll : Maya 2022 505.5 KiB / 2025 506.0 KiB - install copies payload + .mod : both versions OK - uninstall deletes .mod + dir : both versions OK - post-uninstall residue check : zero bytes on both versions - build-missing guard : exit 1 with full cmake recipe Design rationale ---------------- - Maya Module System (rather than dropping .mll into the user's plug-ins dir) gives a clean uninstall surface — one .mod file + one directory, both in a user-owned, non-system path. - onMayaDroppedPythonFile is Autodesk's standard drag-drop hook; stable across Maya 2022 / 2024 / 2025 / 2026. - Per-version plug-in subdirectories are selected at Maya runtime via MAYAVERSION blocks in the .mod file — one install script covers all supported versions, no user choice needed. - Scene node purge (deletes mRBFNode + mRBFShape) precedes unloadPlugin during uninstall; Maya refuses unload otherwise. Stacked on top of Slice 13 — must merge after PR #11. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

d891458249-rgb and others added 6 commits April 21, 2026 20:22

891458249 merged commit 74e8001 into main Apr 21, 2026
4 checks passed

891458249 deleted the slice-13-draw-override-skeleton branch April 21, 2026 21:34

891458249 mentioned this pull request Apr 21, 2026

feat(installer): Maya drag-and-drop installer #12

Merged

8 tasks

891458249 mentioned this pull request Apr 23, 2026

Slice 16 — Phase 2B close-out + v1.2.0 #15

Merged

8 tasks

Navigation Menu

Slice 13 — mRBFShape + mRBFDrawOverride (Path B) - #11

Slice 13 — mRBFShape + mRBFDrawOverride (Path B)#11
891458249 merged 6 commits into
mainfrom
slice-13-draw-override-skeleton

891458249 commented Apr 21, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

891458249 commented Apr 21, 2026

Summary

Path A → B retrospective

Deliverables

Validation status

Invariant changes

Tech-debt added

Test plan

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants