Task: Fix runtime SMR reconstruction to a connected in-game boss body (#72 Phases 1-2)

[jarlbrak]

Summary

Fix the runtime skinned-mesh reconstruction so the Mudwretch Foreman renders as a single connected body in-game instead of an exploded triangle cloud. Apply the minimal fix first (FR-2): after assigning the runtime-built sharedMesh, reassign the live SkinnedMeshRenderer.bones array so Unity rebinds the skin to the new mesh's bindpose set. Only if that does not reach connected:PASS, escalate (FR-3) to a FTK.Stitcher-exact fresh SkinnedMeshRenderer. The two share one acceptance (an in-game connected body on the rigid asset), so they are one deliverable with an internal escalation decided by the gate.

Parent Spec

Implements #72

Responsible Agent

csharp-harmony-engineer

Files to Touch

• FTKModFramework/Core/EnemyVisualPatch.cs (the glb path in ApplyMeshSwap: FR-2 smr.bones rebind; FR-3 fresh-SMR fallback)

Dependencies

• Depends on Task: Skip-skin diagnostic + isolate the boss-shatter to rebind vs decode (#72 Phase 0) #73 (the diagnostic must isolate the failing layer first; if Task: Skip-skin diagnostic + isolate the boss-shatter to rebind vs decode (#72 Phase 0) #73 implicates decode, this work item changes target and the decoder is re-opened before any rebind work)

Acceptance Criteria

• FR-2 (attempted first): after the runtime sharedMesh swap, the patch reassigns smr.bones (and refreshes rootBone only if proven necessary) to force the rebind.
• FR-3 (only if FR-2 does not reach connected:PASS): a fresh SkinnedMeshRenderer is created that shares the live bones/rootBone, the custom mesh and material are assigned to it, and the original renderer is disabled (not destroyed). It is idempotent per combat clone via a name/flag guard, exactly like the existing lantern and golem paths.
• The custom mesh keeps its OWN bindposes (name-remapped to the live rig order, as the loader already does); the live troll bindposes are never reused against the custom vertices.
• The change is confined to Core/EnemyVisualPatch.cs: no public Content.* surface is added and the glTF decoder is not modified.
• It is recorded which fix sufficed (FR-2 minimal rebind, or FR-3 fresh SMR).
• Build green; existing SELF-TEST PASS lines still present.

In-Game Verification

Run tools/ai-model-pipeline/run_visual_gate.py against a bridge-enabled launch (diagnostic flag OFF). The in-game verdict for the Mudwretch Foreman must report connected:PASS: a single connected component, not a shard cloud, on the rigid asset.

A green build alone does NOT satisfy this work item. The in-game connected:PASS is required.

Technical Notes

• In-engine precedent (verified via game-decompile-analyst): FTK.Stitcher is the only runtime skinned-mesh build path in the game (player wearables). It keeps each mesh's own sharedMesh, copies materials, and NAME-remaps bones[] onto the live skeleton via TranslateTransforms; it never sets bindposes/boneWeights/rootBone. Mirror this exactly for FR-3.
• The enemy SMR is otherwise fully prefab-baked (zero runtime enemy bindposes/boneWeights/rootBone assignments assembly-wide), which is why swapping only sharedMesh on the live renderer leaves the skin bound to the previous mesh and scatters the new verts.
• BuildFreshSkin does NOT exist on this branch; FR-3 builds it fresh.
• Do NOT use Mesh.indexFormat / IndexFormat.UInt32 (absent in Unity 2017.2.2p2); 16-bit indices only (21662 verts).
• EnemyVisualPatch re-applies per spawn by design (no _done on the apply); idempotency is PER CLONE.
• If, after BOTH FR-2 and FR-3, the gate still reports connected:FAIL on the rigid asset, do NOT keep grinding: surface it; the spec's objective pivot rule (FR-6) then applies and is recorded in the acceptance work item.

Definition of Done

• Code complete and compiles without errors
• No new SELF-TEST line is expected; existing SELF-TEST PASS lines still appear on startup (no regression)
• In-game run_visual_gate.py reports connected:PASS for the Mudwretch Foreman on the rigid asset
• Recorded which fix sufficed (FR-2 or FR-3); confined to Core/, no Content.* surface, no decoder change

[jarlbrak]

Reconstruction fixed: coherent in-game boss body

The Mudwretch Foreman now renders in-game as a single COHERENT hulking golem (cube head, wizard hat, armored torso, sword), not an exploded triangle cloud. See docs/spike-72-artifacts/AFTER_cleaned_coherent_boss.png.

Root cause was TWO things (FR-1 discriminator + dumps + direct geometry analysis)

1. Rebind (FR-2, the spec's target). The glb path did only smr.sharedMesh = gmesh; and never reassigned smr.bones, so Unity left the skin bound to the previous mesh. FR-2 (reassign smr.bones with a fresh array of the same live bones after the swap, commit 77a9f17) forces the rebind: in-game connected moved from 5 components / fill 0.43 to 2 / 0.84. FR-3 (fresh Stitcher-exact SMR) was NOT needed and would not help (a runtime renderer dump proved a single coherent renderer; customMesh(ftkmf_glb_*)=1).
2. Asset stray triangles (the real residual cloud). The mesh is 100% RIGID (verified in-game RIGID_OK=True, all 17007 verts -> Root_M), so it cannot scatter under skinning. The cloud was ~145 degenerate sliver triangles (1.1%) baked by the offline rig (needles connecting body-center to arm-tips). Removed via tools/ai-model-pipeline/06_clean_slivers.py (commit 15656fc): fracTrisOver10pctDiag 0.011 -> 0.0, rig byte-preserved. The cloud is gone.

The decode path was never at fault (byte-verified by diag_decode_replica.py; offline preview consistently 4/4).

Verification

• Build green; SELF-TEST PASS (x24) with no diagnostic flags; loader bindposes=glb-IBM, dropped slots=0.
• In-game body is a coherent golem in every capture (visual + the RIGID_OK weight dump).
• Recorded which fix sufficed: FR-2 (minimal rebind) + asset sliver-cleanup. FR-3 not required.

Note: a single in-game connected:PASS is observed but NOT stable run-to-run (the silhouette metric is dominated by doorway-vegetation + dark-armor segmentation + combat-pose, count swings 1..5 on the SAME coherent mesh). That gate-stability + the scaled criterion are deferred to epic #65's placement/animation child per FR-4 (recorded in #75). The reconstruction itself (this work item) is fixed and confined to Core/EnemyVisualPatch.cs, no Content.* surface, no decoder change.

[jarlbrak]

Status: PAUSED. The reconstruction half of this task was fixed (smr.bones rebind after the sharedMesh swap; sliver-triangle cleanup of the glb), and those fixes remain in Core/EnemyVisualPatch.cs and tools/ai-model-pipeline/06_clean_slivers.py. But the connected in-game body goal was ultimately met by pivoting the Mudwretch Foreman to the procedural golem body (c6e4b6c), not by the AI glb. See the pause/pivot summary on #72 for the full record. Leaving open: the runtime SMR path was never brought to a gate PASS with the AI glb.