[merynz]

Intent

Independent Mechanics / Motion / runtime generic closure lane; no AXIS/MIRA learning, no IRIS/TESSA/Corpus370 modifications, no Graphics ownership.

Implemented

• Fix real root/subtree bug: ROOT/FK/SOCKET/GLOBAL direct transforms now propagate through the declared joint descendants before constraints, rather than moving only directly targeted joints. Deepest explicit FK overrides local subtree in declared control hierarchy.
• Live Mechanics runtime: same deterministic solver for authored Motion and transient user overrides; exact mesh/rig/skin/Mechanics hashes and constraint schedule; reject undeclared controls and duplicate events.
• Behavioral diagnostics: full LBS deformation, IK residuals and segment strain, typed PASS/FAIL (unreachable targets never report solved), temporal speed/acceleration across user drag inputs.
• Tests: translate whole skeleton/skin; rotated FK child; canonical Motion + mouse override parity; unreachable IK; lineage and negative control tests; deterministic trajectory. CI on Ubuntu Python 3.11.

Scientific boundaries

• Not a source-verified generic mechanics closure; current regression tests use canonical synthetic fixtures. Source interventions remain pending.
• No measured frame-rate claim; no neural runtime calls.
• No joint-limit/DOF enforcement claim, no contact/attachment orientation certification.
• Real Drive Stage47 and V55 evidence refs identified read-only. V55 feature lift is not MIRA skin truth.
• AXIS grammar contains operators unsupported by Mechanics; fail closed and schedule a separate typed expansion court.

Next major phase

• Real source-native Blender intervention vs deterministic Mechanics independent cross-check using hash-exact Drive corpus and protected research-exposed subset.
• Generic DOF/constraint executable semantics, whole-body multi-chain IK with contacts, provenance.
• Temporal continuity and p95 benchmark before product.

Stacked on V6 PR #297; unrelated to active Corpus370 PR #327 and V7 training research #352. No training readiness or model/product authority is minted.

Mechanics continuation (2026-11-06)

• Opt-in simultaneous goals: mechanics_multigoal_solver_v1.py uses declared rotational DOFs + explicitly provided numeric angle limits, joint-tree FK and SciPy least-squares to solve two or more IK goals together, including a planted declared control target. No fabricated joint limits or new control handles. Unsupported pole, DOF, operator and FK-overlay semantics fail closed.
• Live integration: mechanics_live_runtime_v1.py accepts solver_mode="DECLARED_MULTI_GOAL_V1" + pinned_contact_control_ids as an opt-in research mode. Existing product FABRIK remains the default. Unreachable/malformed requests remain visibly FAIL.
• Real source readback: Stage47 truth139_fff1396de9c37b2bab05 Dino source NPZ 24.5 MB and teacher JSON read from Drive, SHA verified; 3,266 vertices / 129 privileged controls / 774 source interventions. Committed research-only, source SHA-locked diagnostic receipt; explicit permanent research-exposure marker (never claim future blind unseen).
• Independent source-integrity code: tools/mechanics/audit_stage47_source_mechanics_v1.py checks manifest SHA, source rest/sham, paired signed actions, root translation and source skin CSR; does NOT claim source→compiled equivalence.
• Courts: multi-goal reachable/unreachable and angular limits; full LBS drag+pin; negative control and Stage47 corruption tests. Separate CI from upstream training.

Remaining hard blockers

1. Qualified source↔canonical joint/vertex maps and independent Blender-vs-Mechanics displacement equivalence on an explicitly research-exposed cohort.
2. Real-valued learned or authored DOF/limits and stable multi-chain contact/grip constraints; current pins are world-position goals, not friction/force contacts.
3. Real-time C++/batched runtime proof; SciPy reference solver not shippable as an assumed 60fps engine.
4. AXIS V6 operator-language/runtime mismatch and source-independent generic morphology tests. No training or product authority minted.

Native source mechanics closure increment (2026-11-06)

• Independent all 774 measured Stage47 action source-local FK/LBS reconstruction against pinned Blender source-world arrays: max vertex L2 error 1.2653386341165218e-6 source units, 774/774 below 2.5e-6 research budget; 43 native bones, 129 isolated 3-axis controls, 3,266 source vertices. Exact source/NPZ hashes in docs/research/MECHANICS_STAGE47_SOURCE_NATIVE_FKLBS_REPLAY_20261106.json.
• Repeatable tools/mechanics/compare_stage47_source_fk_lbs_v1.py includes exact SHA checks, native frame/rotation/skin CSR/parent binding, negative authority cases and optional actual product deform_mesh_lbs_v1 oracle-only probes. CI exercises real product deformation function on independent synthetic source fixture; real Stage47 product-kernel probe is still pending and explicitly distinct from source-local analytical comparison.
• Crucial control-affordance distinction: 5 source bones have no directly skinned vertices, 2 have no skinned descendants, 6 native local rotation parameters have zero measured isolated-response. That does not mean IK pole control role is absent; control function must come from declared constraint graph, not skin weights or privileged source names.
• mechanics_surface_control_effects_v1.py resolves generic declared ROOT/FK subtree and IK/constraint causal effect paths for mouse-picked carrier triangle. Unknown/ambiguous controls abstain or report AMBIGUOUS; no invented handles, no Graphics authority.
• Live multi-goal solver now supports typed bounded warm-start from previous verified frame for motion continuity; retains validated joint-axis identity and no stateful modification of rig/Motion clips.
• Focused Mechanics research CI 32 tests passed on preceding head; full compatibility checks continue as separate workflows.

Scientific firewall: source-native replay is NOT real product output training, NOT qualified canonical source transport, NOT generic unseen success or 60fps certification. IRIS/TESSA, Corpus370 #327 and AXIS–MIRA V7 #352 remain untouched.

Live Mechanics / actual-source product proof (2026-11-06)

• Real product LBS executed on actual private source evidence: independently downloaded and SHA-checked Stage47 Dino input. 128/128 stratified source intervention samples passed original deform_mesh_lbs_v1 vs Blender and prepared kernel vs both. Maximum original→prepared vertex L2 difference 8.95090418262362e-16 source units; original vs Blender max 9.487445122129821e-7. Full 774-source-action analytical reference passes, but all-774 product kernel was not completed.
• Measured performance, no FPS claim: 128-action sample, full 3,266-vertex source mesh, 6,571 source weights and 43 native bones; median LBS-only CPU 66.0ms → 3.61ms, 18.28× observed speedup, one-time preparation 4.25ms on one container. Does not include solve, Graphics, render or p95.
• Opt-in cached sparse skin: mechanics_prepared_skin_v1.py pre-packs immutable CSR-style influence coordinates and uses vectorized NumPy transformation; mechanics_live_runtime_v1.py can consume prepared_skin=, default legacy unchanged. Full parity + negative tests.
• Control picking authority repair: mechanics_surface_control_effects_v1.py verifies actual CanonicalMesh face/topology and executable rig constraints; orphan IK and unsupported operators fail closed. mechanics_control_response_court_v1.py independently runs small ± control perturbations through real Mechanics + LBS and distinguishes graph-potential from observed local effect; never infers role from source bone labels.
• Evidence: docs/research/MECHANICS_STAGE47_PRODUCT_LBS_PREPARED_128_20261106.json and appended docs/research/MECHANICS_GENERIC_CLOSURE_V1_20261106.md. All producer evidence remains read-only; artist source bytes not committed; temp code-only artifact uploader removed after experiment.
• Scientific gap: real upstream source→canonical geometry/rig/skin binding remains unqualified. Neither generic unseen, learned control, training nor product-release authority minted.

2026-11-06 complete Mechanics CPU-frame optimization

• Runtime court fixed without evidence loss: vectorized full mesh vertex displacement/finite check and cached immutable rest positions; still verifies every vertex and joint, exact IDs and lineage. Negative cases cover vertex reorder, duplicate IDs, nonfinite data, stale cache.
• Executable CPU frame benchmark tools/mechanics/benchmark_live_mechanics_v1.py: actual 2-chain/2-IK, one planted target, warm-start and full live Mechanics + LBS and behavioral diagnostic on seeded synthetic carrier at 3,266 and 8,192 vertices.
• GitHub Actions verified on commit dcefe18c5b7efc3b55eef1076ef14c9bbcc2bfb2, run 42273795156: 48 unit/integration tests passed. At 3,266 vertices, full Mechanics P50/P95 5.924/6.393 ms vs sampled legacy P50 57.807 ms; at 8,192, P50/P95 11.351/12.109 ms vs legacy P50 143.126 ms. Max prepared-vs-legacy vertex L2 difference 4.48e-16.
• Reproducibility receipt: docs/research/MECHANICS_CPU_FULL_FRAME_COURT_20261106.json; Actions benchmark diagnostic artifact id 13531142924 (7-day retention).
• Critical limit: only 5 joints / 3 rotational DOFs, small samples, host-specific CPU; graphics, real input, GPU upload, collision and force-based contact excluded. This does not certify product FPS. Next: DOF-scaling and explicit Jacobian/morphology reference tests, then source↔canonical acceptance.
• No changes to IRIS/TESSA, Graphics compiler, Corpus370 truth, or AXIS/MIRA training pipelines.

Analytic DOF-scaling increment (2026-11-06)

• Added explicit opt-in ANALYTIC_RESEARCH_V1 rotational FK Jacobian (default NUMERIC_V1 unchanged), preserved declared DOF/angle limits and pinned world-space IK residual semantics.
• Central finite-difference independent gradient court and reachability/negative cases: 3, 6, 9, 15 DOF, 55 focused tests PASS on GitHub Actions run 42275027914, tested head 315d13e.
• CPU solver-only p50, numeric→analytic: 3 DOF 4.194→3.635 ms, 6 DOF 6.917→4.763 ms, 9 DOF 9.776→5.833 ms, 15 DOF 15.445→7.576 ms (2.04×). Gradient max differential error ~1.24e-9. This is a 5-joint synthetic numerics fixture, not a source-real morphology or product FPS proof.
• Reproducibility: tools/mechanics/benchmark_multigoal_jacobian_v1.py and docs/research/MECHANICS_ANALYTIC_JACOBIAN_DOF_SCALE_20261106.json.
• Research remains isolated from IRIS/TESSA, Graphics relations, Corpus370 and AXIS/MIRA training. Next gate is longer generic trees, source↔canonical frame/DOF authority, simultaneous contact/grip orientation, verified complete product latency.

Longer articulated trees, world contact anchors and explicit actuator policy (2026-11-06)

• Actual 24/48/72 pivot DOF trees (27/53/79 joints), curved branches in XY/XZ/YZ, declared IK controls on 2/4/6 tips, simultaneous 1/2/3 world-position contact pins and 1/2/3 dragged tips. No source bone names, teacher bytes or model inference.
• Acquired world pin support: new pinned_world_targets argument bound exactly to declared pins, preserved through live Motion/drag. Bad/missing/nonfinite anchors and pin/override conflict fail closed.
• Observed consequence vs representation authority: all 24/48/72 rotational DOF columns individually affect a tip at rest, but endpoint-only Jacobian rank is only 4/8/12. The 20/40/60 null dimensions are NOT thereby source-disproved; independent motion/skin/contact evidence is needed.
• Full solver reached all goals with numerical segment-length preservation. GitHub runner single solve timings 94.01/277.44/422.36 ms — too slow for shipping real-time multi-DOF.
• Explicit optional active_parameter_keys: freezes unselected declared DOFs at their verified warm-start values, leaves CanonicalRig intact and still checks every original goal residual. On synthetic generator-known one-pivot targets, oracle-subset 1/2/3 active axes solved in 8.51/26.42/39.20 ms (~10–11x observed). A generated two-pivot target fails the one-pivot policy and succeeds in full mode. Selection is a representation/runtime policy, not source truth and not generic automatic selection.
• Verification: workflow run 42278009295 on code SHA 2e0cccf89828243133fbdcb45189d36f4ea565b5 reports 68 tests PASS; reproducible receipt docs/research/MECHANICS_LONG_BRANCHING_DOF_POLICY_20261106.json, scenario generator tools/mechanics/benchmark_long_branching_ik_v1.py.
• Next hard problem: independent task-space rank-revealing active subspace selection with deterministic fallback for unsatisfied contact/drag and sequence continuity; later, source-native intervention correspondence and position+orientation/attachment tests. No merge or training authority minted.

First generic task-space actuator decision candidate (2026-11-06)

• Opt-in non-oracle RANK_REVEALING_RESEARCH_V1: derive exact world-space goal Jacobian from declared CanonicalRig, exclude regularization rows, select independent actuator columns with pivoted QR; freeze unused declared DOFs at validated warm-start, check every original constraint and retry all DOFs if reduced solve fails. No source truth labels or control names inform selection.
• Morphology scaling verified: 24→4 selected, 48→8 selected, 72→12 selected DOFs on curved 2/4/6-branch synthetic trees with matching 1/2/3 world position pins and dragged control goals. Full vs rank-policy one-shot CPU solving: 74.08→9.90 ms (7.48x), 228.54→28.27 ms (8.08x), 351.21→42.43 ms (8.28x). No lost goal constraints in these cases.
• Negative safeguard: at a singular straight rest pose, full-DOF fallback is actually invoked but the local full solver still FAILs the nonlinear target; the result remains FAIL, never fake PASS. This is an open nonlinear optimizer and temporal continuity gap.
• GitHub CI: 73 focused tests PASS, run 42279339544, tested SHA a9abd29750be4ca84fde2ef9ed55b1ce7fa8af67. Source-of-truth for data/limitations: docs/research/MECHANICS_RANK_REVEALING_ACTUATORS_20261106.json.
• Scientific boundary: only synthetic task-space results, not learned AXIS/MIRA, native authored DOF, real character correspondence, stable motion quality or whole-product 60 FPS. Rank policy is a representation/runtime policy, not source truth or a change to the puppet capability contract.

2026-11-06 Spine-informed (independent) temporal Mechanics research

• Secondary after primary constraints: declared trailing 3D joint chains only; critically damped angular-bounded motion; no overlap with authored IK/contact/attachment ownership. Parent-child integration occurs per internal physics substep, preserving tested 1×1/30 vs 2×1/60 fixed-target equivalence within 1e-11. Policy changes require explicit reset via SHA-bound state. No teacher-derived physical stiffness or force/collision claim.
• Rank-selection hysteresis: reuse a previously independently spanning actuator basis if the current task Jacobian still supports it. Separate real rank/span changes from unjustified control flicker; test shows zero unjustified churn across seven dragged frames.
• Bounded singularity recovery: opt-in up to 12 reproducible restarts on declared DOFs only; an analytically witnessed reachable straight-chain radial target is rescued, while an impossible pinned target retains FAIL. No global nonlinear convergence promise.
• 3D attachment world orientation evaluator: declared local grip basis + primary joint solved world rotation/position; missing source basis is explicitly unqualified; wrong required orientation FAIL. Does not solve oriented IK or invent a grip control.
• Spine is a runtime design reference, not equivalent engine or copied code: consulted public constraint scheduling, physics FPS/reset and transform docs; no runtime SDK sources imported.
• Verified focused CI: 94 tests PASS and Mechanics ownership firewall PASS at SHA 668bbd44eaa3d26898be8864fe1bfb198587c44b, Actions run 42284413119. See docs/research/MECHANICS_TEMPORAL_SECONDARY_ATTACHMENT_20261106.json, docs/research/MECHANICS_SPINE_RUNTIME_INDEPENDENT_DESIGN_20261106.md.
• Still open: true orientation IK, physics/collision/contact forces, generalized source→canonical morphology proof, arbitrary scrub/replay and full rendered 60fps. IRIS/TESSA, Graphics authority, Corpus370 and AXIS/MIRA training lanes untouched.

2026-11-06 — Oriented and full 6D grip IK: verified independent Mechanics result

• New opt-in analytic multi-goal attachment frame solver: declared anchor position local_P and optionally authored metadata.local_basis become 3D world grip-position plus 3D world grip-orientation objectives, combined with existing drag and pinned-contact residuals.
• Exact FK gradients: grip position derivative axis × (worldGrip - pivot); 9-row world-basis chordal rotation derivative axis × worldBasisColumn. Numeric centered-difference regression independently checks these derivatives. Final orientation qualification is in radians from SO(3) geodesic, not optimizer termination.
• Temporal execution: MechanicsInterventionStepV1 supports per-frame changing grip position and world orientation; continuous five-frame synthetic sequence retains fixed opposing contact and warm start. Conflicting static+keyframed authorship, missing grip basis, unknown anchors, incompatible position/orientation all fail closed.
• Compatibility: EVALUATE_ONLY remains default, SOLVE_DECLARED_RESEARCH_V1 requires opt-in analytic multigoal mode. Neither CanonicalRig authority nor Graphics/CAA, Corpus370, IRIS/TESSA, AXIS/MIRA training changed.
• Verified focused GitHub Actions: 106/106 tests PASS, runtime research firewall PASS at tested code head d76362eb9d5970a76aee8067b9ec300d4d631738, workflow run 42286207240. Reproducible receipt docs/research/MECHANICS_DECLARED_6D_GRIP_IK_20261106.json.
• Not yet proven: independently sourced real grip geometry/frame, real 43-bone+ poses, interaction with multiple oriented tools, collision/friction, entire product FPS or unseen model learning. Research-only.

2026-11-06 — Multiple rigid grips: independent conflict witness and explicit priority (research-only)

• New mechanics_attachment_feasibility_v1.py computes necessary same-parent rigid-frame feasibility for multiple world position/orientation targets. Emits deterministic witness kinds RIGID_POINT_DISTANCE, RIGID_PARENT_ORIENTATION and RIGID_FRAME_POINT_DISPLACEMENT, including target mismatch, tolerance, parent and anchor IDs. Lack of witness is UNKNOWN global IK feasibility, never a PASS.
• Optional solver attachment_preflight_mode="RIGID_HARD_NECESSARY_V1" fails closed before optimization on a proven hard same-parent contradiction. Default previous behavior remains unchanged.
• New mechanics_attachment_priority_v1.py provides strictly explicit per-track HARD / ADVISORY application policy; all IK controls and world contact pins always remain HARD. Advisory goals are not discarded: independent attachment frame evaluation measures each residual. A result can report PASS_HARD_WITH_UNMET_ADVISORY_TARGETS while all_requested_targets_satisfied=False; it cannot be conflated with full compile PASS. No automatic priority derivation or source authority inferred.
• 14 new focused positive/negative tests across both modules; GitHub Actions run 42288736079, tested code head 36e69a1fda55bf83944e74502e3a3ea7fd67d988: 120 Mechanics tests PASS, ownership firewall PASS. Evidence: docs/research/MECHANICS_MULTI_ATTACHMENT_PRIORITY_FEASIBILITY_20261106.json.
• Open: cross-parent nonlinear feasibility, source-authored grip evidence/priority authority, temporal advisory-policy switches, real source→canonical correspondence, GPU/render throughput. No changes to Graphics/CAA, Corpus370, or training.

2026-11-06 — Explicit cross-parent shared-rigid-prop 6D grip and grounded third limb

• New research-only declaration/court mechanics_relative_attachment_bridge_v1.py: SHA-bound named pair of attachment anchors on different parent joints, each with explicitly declared local grip frame, plus declared 3D relative position and 3D relative rotation. No automatic same-weapon detection, inferred source grip authority or physics/stiffness truth.
• Exact analytic FK integration: combined 3+9 relative grip residual rows and Jacobian, with existing explicit drag, 6D absolute grip targets and world contact pins. Rank/recovery policy sees these rows. Full solver residual and independent post-primary world-frame bridge court both run; failing bridge cannot be hidden by secondary dynamics.
• Behavioral test: a single A branch drag moves B branch without a direct B control due to the declared link; contradiction between a planted B goal and A drag fails rather than reshaping rigid link. A more realistic synthetic three-branch scenario moves a shared two-grip prop over five target angles while a separate third distal limb maintains its fixed world contact, including warm starts and skin deformation.
• GitHub verified: run 42291808979 at code SHA a240f1ba08cc2c01ff6718ef0edbab65fe034601: 128/128 Mechanics tests PASS, research ownership firewall PASS.
• Measured diagnostic CPU benchmark from same run: 12 measured frames + 3 warmup, 8 joint / 5 rotational DOF / 3 skin vertex synthetic rig: Mechanics full-frame p50 4.129 ms, p95 4.199 ms; max world contact error 4.386e-5, relative grip position 2.193e-5, relative grip angular 1.096e-5 rad. This is emphatically NOT full-product 60fps or source-native character evidence.
• Reproduction/receipt: tools/mechanics/benchmark_relative_attachment_bridge_v1.py, docs/research/MECHANICS_CROSS_PARENT_6D_GRIP_GROUNDED_THIRD_LIMB_20261106.json.
• Remaining gates: source-authored prop attachment and correct local grip reference frames, generic unseen morphology and independent source↔canonical consequence, real collision/friction, larger model meshes, end-to-end Graphics/CAA frame budget. Existing Graphics/Corpus370/IRIS/TESSA/AXIS-MIRA training lanes untouched.

2026-11-06 — Post-V7 executable puppet product gap assessment

Critical interpretation: Mechanics research is real code, but cannot yet be called V7-native shipping product or a compiled generic puppet. V7 PR #352 is a separate draft, using the V6 grammar. PR #353 is also a draft sibling on the same V6 base; they have not been combined into a source-bound ModelRelease/Graphics/Editor/export proof.

• New machine-readable static gap audit: tools/mechanics/audit_v7_mechanics_execution_gap_v1.py AST-reads actual V6 producer vocabulary and Mechanics executor definitions without importing Torch or inventing a release. 8 vs 5 declared/ executable operators; unsupported in the current canonical operator schema: LIMIT_ROTATION, LIMIT_TRANSLATION, PATH. 9 vs 3 V6 local DOF labels vs analytic multigoal axes; no separate typed ANCHOR 3D grip frame or two-anchor relative 6D bridge payload in the event contract. These are contract gaps, NOT automatic release errors concealed by the compiler: source/code generation and downstream fail-closed gates need individual qualification.
• Verified Actions run 42292882374: 130/130 Mechanics tests PASS, ownership firewall PASS; typed V7 static compatibility verdict stays explicitly GAP__NO_V7_MECHANICS_PRODUCT_READINESS. Negative regression rejects stale / unrecognized source constants.
• Product gap map: docs/research/MECHANICS_POST_V7_PRODUCT_GAP_MAP_20261106.md. P0 is cross-layer vocabulary/DOFs, authored local grip + relative bridge, editable whole-body root and independent source↔canonical transfer, V7 model/Corpus370 admission, plus sibling-branch product integration. P1 is contact/manifold, proper joint limits and constraint dependencies, multiple collisions, physical secondary behavior and temporal replay. P2 is realistic-frame runtime, read-only Graphics handoff, editor/clip/export and verified proof.
• Do not merge or claim release on 130 synthetic tests: continue with typed Mechanics capability manifest and complete declarative V7 program execution coverage first. No source, graphics, training or learning authority changed.