Task: Full 4-criterion gate PASS, close divergence, remove diagnostic, record B-or-pivot (#72 Phases 3-4)

[jarlbrak]

Summary

Take the now-coherent boss body to the spec's Definition of Done: a full four-criterion in-game gate PASS for the Mudwretch Foreman with the offline and in-game verdicts agreeing (the offline-PASS / in-game-FAIL divergence closed), then remove or fence the FR-1 diagnostic toggle, and record the objective B-robust-or-pivot decision (FR-6) backed by the controlling gate output.

Parent Spec

Implements #72

Responsible Agent

csharp-harmony-engineer

Files to Touch

• FTKModFramework/Core/RuntimeGltfMeshLoader.cs (remove or clearly fence the FR-1 diagnostic toggle)
• ninum hub (proj_4f015453) and the epic Spec: Runtime-glTF approach spike (B): coherent in-game boss body, gated to PASS or pivot #72 / Epic: Robust custom enemy/boss 3D-model pipeline (editor-free, in-game verified) #65 threads (record the FR-6 decision with gate evidence)
• No gate file is edited (run_visual_gate.py, visual_gate.py, visual_verdict.py, the harness driver, baseline.json, thresholds are reused as-is)

Dependencies

• Depends on Task: Fix runtime SMR reconstruction to a connected in-game boss body (#72 Phases 1-2) #74 (a coherent connected body must exist first)

Acceptance Criteria

• run_visual_gate.py reports an in-game four-criterion PASS (connected, upright, centered, scaled) for the Mudwretch Foreman.
• The same run prints the OFFLINE preview verdict and the IN-GAME verdict, and they AGREE on the four mechanical criteria (the divergence is closed). Any residual disagreement is proven to be offline-preview framing only (a normalized-metric artifact of the approximate crypt camera), NOT a body-coherence difference, and is logged as a gate-tooling follow-up under epic Epic: Robust custom enemy/boss 3D-model pipeline (editor-free, in-game verified) #65's later "Placement, scale, and animation correctness" child; it does NOT block the B-robust verdict.
• The FR-1 diagnostic toggle is removed or clearly fenced as diagnostic-only; behavior is byte-identical with it off and SELF-TEST PASS holds with the flag off.
• No gate file or threshold is modified to obtain the PASS.
• The FR-6 decision is recorded with the controlling gate output: either "runtime-glTF (B) is robust editor-free", or (if the reconstruction could not reach connected:PASS in Task: Fix runtime SMR reconstruction to a connected in-game boss body (#72 Phases 1-2) #74 even after the fresh-SMR escalation) "B is not robust editor-free; pivot to A (AssetBundle) or C (procedural)", naming which approach the next spike should pursue and why.
• The decision and evidence are recorded in ninum (proj_4f015453) and summarized on the epic Epic: Robust custom enemy/boss 3D-model pipeline (editor-free, in-game verified) #65 thread.

In-Game Verification

Run tools/ai-model-pipeline/run_visual_gate.py (diagnostic flag OFF) against a bridge-enabled launch. It must print an in-game four-criterion mechanical PASS for the Mudwretch Foreman AND show the offline and in-game verdicts agreeing on those four criteria. With the toggle removed/fenced, the framework still logs the existing SELF-TEST PASS lines.

A green build alone does NOT satisfy this work item. The 4/4 in-game PASS plus offline/in-game agreement is required (or, on the pivot branch, the recorded objective pivot verdict with the controlling gate output).

Technical Notes

• "Divergence closure" means the two verdicts agree on the four BOOLEAN criteria (the in-game connected:FAIL that signaled the runtime shatter is gone). NFR-4 forbids editing the gate to reach this; the only sanctioned escape hatch is the offline-framing-only follow-up described in the acceptance criteria.
• The scale band is centered on ~1.4 (BossBodyScale = 1.4 vs the scale-1 stock baseline), already finalized in spec Spec: Visual-verification gate (objective in-game PASS/FAIL for custom enemy bodies) #66's thresholds. Do not retune thresholds here.
• The B-robust verdict rests on the rigid-asset PASS; optional multi-bone polish is a separate work item and is not required for this one.

Definition of Done

• Code complete and compiles without errors
• No new SELF-TEST line is expected; existing SELF-TEST PASS lines still appear with the diagnostic flag off (no regression)
• In-game run_visual_gate.py prints 4/4 mechanical PASS + offline/in-game agreement for the Mudwretch Foreman
• FR-1 diagnostic toggle removed or clearly fenced; no gate file modified
• FR-6 decision (B-robust or pivot) recorded with gate evidence in ninum and on epic Epic: Robust custom enemy/boss 3D-model pipeline (editor-free, in-game verified) #65

[jarlbrak]

FR-6 decision recorded: runtime-glTF (B) is ROBUST for reconstruction; diagnostics removed

B-robust verdict (FR-6)

The runtime skinned-mesh reconstruction was fixed: the boss renders as a coherent hulking golem editor-free (see #74). B can build a coherent custom boss body at runtime; no pivot to A (AssetBundle) or C (procedural) is warranted for the reconstruction problem. The decode/rebind/loader path is sound.

Full 4/4 in-game PASS was NOT reached; recorded as a deferral (FR-4 escape hatch), not a B failure

The user reviewed the evidence and accepted closing on the reconstruction win, deferring the rest. The two blockers are orthogonal to the A/B/C reconstruction approach and were proven NOT to be coherence defects:

1. scaled is structurally unreachable for this creature. The band (height_ratio [1.20,1.65], area_ratio [1.45,2.60]) is calibrated for the BULKY stock troll; the slim AI golem tops out around area_ratio ~1.0 even fully detected at proper height. NFR-4 forbids relaxing the gate.
2. In-game connected/upright are scene/pose dominated, not coherence. On the SAME final config the component count swung 1, 2, 3, 5 and upright 7deg..65deg, while the mesh is provably constant (rigid, cleaned, RIGID_OK=True, offline 4/4 every run) and every crop shows the same coherent golem. The variance is doorway VEGETATION segmented as extra components + dark-armor noise + combat animation pose at capture.

Both are filed for epic #65's "Placement, scale, and animation correctness" child + gate hardening (fixed-pose capture / boss-region masking; in-place feet-anchored scaling; per-chassis scaled calibration). NFR-4 honored: NO gate file or threshold was modified.

Diagnostic hygiene (FR-5 / NFR-5): done

All temporary instrumentation removed (commit 823a6dd, verified zero remaining): FTK_DIAG_SKIP_SKIN, the WDIAG built-weight dump, the renderer dump, and the FTK_BOSS_SCALE/WIDTH/EMIT env levers. RuntimeGltfMeshLoader.cs decoder net change is zero. SELF-TEST PASS (x24) holds with no flags.

ftk-architect review: addressed

CRITICAL finding fixed (commit 6b95ebd): a Core-side baked scale that compounded with the registered EnemyVisual.scale (~2.8x) and was invisible to the self-test was removed; the boss is scaled by the single registered-scale authority. Boundary clean (no Content.* surface, no FTK_*DB, no new ids), net35-safe, decoder untouched.

Evidence

docs/spike-72-artifacts/: SPIKE-72-FINDINGS.md (full writeup), BEFORE/AFTER crops, phase-0 discriminator crops, the original troubleshoot brief. Controlling gate output across runs recorded therein.

Final state: build green, SELF-TEST PASS, coherent boss, all diagnostics removed, single scale authority. Commits 80fb4fd..dd61ef0 on campaign-d1-hollow-mire.

[jarlbrak]

Status: PAUSED, with the B-or-pivot decision effectively made: PIVOT. The shipped Mudwretch Foreman uses the procedural golem body (c6e4b6c); the AI-glb static-attach experiment was removed (45dd6dc). The full 4-criterion gate PASS was never reached for the AI glb (scaled band is troll-calibrated, connected/upright are scene- and pose-confounded per SPIKE-72-FINDINGS.md). Remaining before this can close: record the corrected verdict in SPIKE-72-FINDINGS.md, commit the CORRECTION/RENDER/MESHQUALITY evidence artifacts, and close the divergence per the task title. See the fuller status on #72; resume notes in ninum kb_fbc9fbde.