Project TX video extractions into hugs research; dedupe AGENTS.md grab item; close research phase

AGENTS.md

@@ -210,20 +210,14 @@ device only; the editor has an XR simulator for dev. Trackers/action map: see `x
    snap — SUSPENDED for a hand while it grips, HAND ghosts > 10 cm lag). Collision layers:
    1 = world, 2 = UI (panel mask 0 — props pass through even when shown); layers live on
    CollisionObject3D, NOT CollisionShape3D. Panel = prop-UI prototype (ax_button summon,
-   contact-press knobs, Reset). Research reports live in `research/` (arm IK, IK
-   stability/deep, grabbing, two-handed grab/Box3D internals, addon weld patch) — READ before
-   re-researching or changing grab/IK. Stink-O's box3d-godot (custom components) = documented
-   plan-B. Laser/distant grab REMOVED (user never knew it existed); grab = palm-proximity only,
-   per-hand palm-face offset (left +X, right −X — taste). Box3D PinJoint quirks: anchors are
-   CAPTURED AT CREATION; the joint node's position is READ EVERY TICK (keep it on the palm =
-   hand-following pivot); break must measure the prop-local anchor vs the palm point. Headless
-   regression tests: `scenes/physics_spike.tscn` AND `scenes/grab_spike.tscn` — both must print
-   overall PASS. `PhysicsHands.cs`: physical palm bodies (0.4 kg, 90 N chase, rotation lock,
-   stuck-safety snap, HAND ghosts > 10 cm lag). Collision layers: 1 = world, 2 = UI (panel mask
-   0 — props pass through even when shown); layers live on CollisionObject3D, NOT
-   CollisionShape3D. Panel = prop-UI prototype (ax_button summon, contact-press knobs, Reset).
-   Research reports live in `research/` (arm IK, IK stability/deep, grabbing,
-   two-handed grab/Box3D joint internals) — READ before re-researching or changing grab/IK.
+   contact-press knobs, Reset). Laser/distant grab REMOVED (user never knew it existed);
+   grab = palm-proximity only, per-hand palm-face offset (left +X, right −X — taste).
+   PinJoint quirks (anchors captured at creation; joint node position READ every tick;
+   break measures the prop-local anchor vs the palm point) apply to the v5.0–5.2 PIN path
+   only — the weld does not share the pivot behavior. Research reports live in `research/`
+   (arm IK, IK stability/deep, grabbing, two-handed grab/Box3D internals, addon weld patch,
+   body physics hugs, body physics Godot facts) — READ before re-researching or changing
+   grab/IK/body. Stink-O's box3d-godot (custom components) = documented plan-B.
 7. **Getting a body (rig) — placeholder v1.2 (cube character, agent-built).** `BodyRig.cs`:
    all-BOX primitives (owner preference) — head 0.15³ behind the eyes, slim torso retracted,
    wrist-anchored 2-bone arms. **Soft IK applied**: shoulder→hand distance clamped to

HANDOFF.md

@@ -263,11 +263,16 @@ PropShop, PropGrabber, PhysicsHands, BodyRig, FadeLayer/Fade. Action map: `openx
    wants to pursue a full physics VR body (also as reusable tech for a second project idea);
    the hug should be mostly physics with detection for future interactivity. Research done:
    `research/vr-body-physics-hugs.md` (prior art + phased design; §6 = Nimso's "Hybrid
-   Presence", the closest known implementation, closed-source, only videos) and
-   `research/vr-body-physics-godot.md` (verified stack facts: no Generic6DOF, hinge velocity
-   motors work, PD-follow proven by the hands). Owner offered video-extraction of Nimso's
-   deep-dive videos — pending. Recommendation shape: staged torso → 2-segment arm chains
-   (chase-driven, Boneworks/Nimso tier, NOT per-bone skeleton sim), test hugging props.
+   Presence", the closest known implementation; §7 = owner-extracted video breakdowns of
+   Project TX: hierarchical per-joint force budgets, script-relaxation hand adaptation,
+   active-ragdoll NPCs with grab anchors) and `research/vr-body-physics-godot.md` (verified
+   stack facts: no Generic6DOF, hinge velocity motors work, PD-follow proven by the hands).
+   Research verdict: the technique is an independent-implementation match of our existing
+   chase-drive architecture; no more video extraction needed now (conditional: Spideriini
+   elbow handling if Stage-2 elbow jitter bites; Test Chamber if a locomotion stage ever
+   starts). Recommendation shape: staged torso → 2-segment arm chains (chase-driven,
+   Boneworks/Nimso tier, NOT per-bone skeleton sim), test hugging props; muscle model =
+   hierarchical budgets (hand ≺ forearm ≺ shoulder ≺ torso) scaled from body mass.
    Character model: keep boxes for the physics work (owner asked; see research §5-6).
 2. **Grab points on special objects** — grab is palm-proximity to the prop origin today;
    special objects may want authored grab anchors (see `research/vr-grabbing.md` taxonomy).

research/vr-body-physics-hugs.md

@@ -187,3 +187,55 @@ no written technical breakdown exists anywhere.
   force-clamped chase drives, i.e. the same pattern as our PhysicsHands generalized to limbs;
   "hybrid" = tracked targets + simulated bodies, which is ALSO our layer-3 architecture
   (AGENTS.md). We are not inventing a new technique — we'd be porting a known-shaped one.
+
+## 7. Project TX video extractions (owner-provided AI summaries, 2026-10-01)
+
+Owner extracted AI breakdowns of Nimso's videos: "IN-DEPTH VR PHYSICS INTERACTIONS" pt. 1–2
+(`youtube.com/watch?v=tMDmwD1cQPg`, `BTlY8SWLtgo`), plus asks on joint/force model, muscle
+system, hand adaptation, and physics NPCs. Distilled to what bears on OUR design:
+
+- [S] **Force model = ours, generalized.** Script = "brain" reading tracker data, feeding
+  targets; solver executes via ConfigurableJoints in force/acceleration mode with PD-tuned
+  spring (Slerp) drives — never infinite force. Heavy object ⇒ spring stretches ⇒ hand lags
+  the controller ⇒ weight feel. Exactly our PhysicsHands pattern (90 N chase) and the grab
+  weld's philosophy: the solver owns the coupling, the script owns the intent.
+- [S] **Muscle = hierarchical per-joint force budgets.** Base avatar mass 60 kg scales the
+  joint drives' F_max. Shoulders get HIGH torque budgets (climb-hangs the whole body);
+  wrists/forearms get strict caps. A 12 kg sledgehammer held out straight exceeds the
+  wrist/forearm cap via lever arm ⇒ hand buckles; two-hand sliding down the shaft shortens
+  the lever ⇒ liftable. Fatigue is modeled as mechanical disadvantage ONLY — no stamina
+  meter; strain reads through visual lag and drift. → Refinement for our Stage 2: replace
+  the flat 90 N hand budget with per-joint budgets (hand ≺ forearm ≺ shoulder ≺ torso)
+  scaled from a body-mass stat; props with attachment coordinates register their mass at
+  the attachment point (holstered 12 kg hammer scaled the whole avatar 60→72 kg).
+- [S] **Hand adaptation = script-side relaxation on contact (anti-winch).** When the hand
+  hits structural limits, the script temporarily relinquishes axes, SHIFTS THE PIVOT (palm
+  → fingertips against a magazine baseplate), and eases joint boundaries dynamically
+  ("dynamic radius gripping" tracks a bat's changing radius as the hand slides). The winch
+  lesson in reverse: don't fight the solver — reconfigure. Our analog already exists as
+  break thresholds; a future nicety is pivot-shift on contact for resting palms on surfaces.
+- [S] **Physics NPCs (the huggee half) = active ragdoll + kinematic grab anchors.** The NPC
+  is a live ragdoll driven by internal joint-matching springs (Boneworks-class). Strikes
+  compare impact velocity against NPC joint stiffness; enough momentum overrides the NPC's
+  balance solver (deflection before hit registration). Grabbing an NPC's hand/head creates
+  a TEMPORARY fixed/distance joint between the player's hand rigidbody and the NPC
+  hitboxes — the NPC instantly feels the player's weight; you can swing, push off-balance,
+  pull close. → For us: hug/touch = contact between player chest/palms and an NPC
+  active-ragdoll; grabbing NPCs = our existing ConeTwist weld against NPC anchor bodies;
+  squeeze-back = NPC joint-matching drives ramping against the player's budget.
+- [S] **Leg colliders exist but are buggy** (legs lock up, impede crawling) — the dev
+  shipped them anyway for climbing/tire-swing play. Validates deferring legs LAST.
+- [S] **Locomotion is physics-native**: stick push at slow speed, natural sway raises walk,
+  jog bounce frequency triggers run ("remember your feet aren't tracked"); two-handed
+  weapons blend primary-hand rotation with secondary-hand position ("Hybrid system").
+- [S] **Out of scope for VR Messs** (noted for completeness): extended physics materials as
+  data wrappers (surface-pair impact audio, density-based stabbing, material magnetism),
+  dynamic recoil from muzzle/bullet-mass integration, geometric mag-well sliding with
+  ergonomic posing rules, mass-independent buoyancy (surface-area/displacement forces so a
+  1 kg box floats a 70 kg body without jitter), modular attachments shifting center of mass
+  locally, cross-compatible ammo systems.
+- [I] **What the extractions change in our plan:** confidence only + two concrete design
+  upgrades (hierarchical budgets; NPC anchors via existing welds). No architectural
+  surprises — an independent implementation converged on our exact shape. Remaining unknown
+  = tuning numbers (per-joint gains; published nowhere, must be tuned in-headset) and
+  locomotion/balance/comfort (only relevant at a Stage 3+ we may never reach).