Draft: V0 Animation Correction — Loops 1–6 (do not merge)

[HSpector1]

Draft checkpoint — DO NOT MERGE. Animation-led visual correction of the V0 3D restaurant
readability sandbox. Hybrid characters = Quaternius Universal Animation Library (CC0, 65-bone rig,
43 clips) + custom identity skinned/bone-parented to the same armature. Raw UAL mesh git-ignored;
only customized derivative GLBs are committed.

Status

• 6 of a minimum 10 loops complete. Target 12 loops — NOT complete.
• Human testing is NOT authorized and has not been run.
• Do not merge.
• Backend integration and M1 are NOT authorized and not begun. No .cs/src/tests/M0 touched.

Loops

• 1–3 superseded procedural foundation (retained as evidence).
• 4 adopt hybrid; server/cook identity. Pass-with-notes.
• 5 customer cast + interaction; targeted remediation → all gates pass (prop 4.3, action 4.2, style 4.4).
• 6 — Host & Busser operational language (this commit, 9d2f07e): host + busser characters, animation
  language, physical table-state props, integrated healthy/overload host+table-turn scene with
  presentation/readability HUD layers.

Loop 6 verdict: Fail — targeted remediation (Loop 7 not authorized)

Foundation is sound: characters/clips build, no regression in Loop 5 work, pipeline reproducibility
4.5, Godot reliability 4.5, deformation 4.0, performance honest (10-char cast 166k tris / 590 draw calls;
LOD1 −24.8%). But the primary gate fails — the entrance/table-turn state does not read at the default
management overview without the HUD: the physical table carries no state that survives zoom-out or
greyscale, and host/busser reads are prop-dependent. Six categories fall below the critical-3 floor
(dirty-table 1.5, busser/host at overview 1.5/2.0, host-turnover-failure 2.0, busser-action 2.5,
table-turn 2.5). Scores from four independent reviews; not inflated.

Full evidence: reports/visual/v0-animation-correction/iteration-06/ (HYPOTHESIS, CRITIQUE with 4
independent reviews, SCORES, DECISION, lineups, reels, sequences, state lineup + greyscale, prop
closeups, 4 overview recordings, performance-counters.csv, validator.txt, regression-checklist,
loop5-vs-loop6). Recommended next: targeted remediation of table-state-at-overview + host/busser identity.

[HSpector1]

🔬 Foundation Decision Spike — result (clearly-labeled spike evidence; not Loop 4)

Bounded procedural-vs-Quaternius-vs-hybrid comparison, on branch spike/v0-foundation-comparison (from this PR's head 07b633a). PR #8 not modified/merged. No Loop 4. No manual weight-painting performed.

Decision: Adopt HYBRID deformation foundation (Quaternius CC0 rig + custom restaurant identity)
Verdict:  Conditional-Pass
Action:   Authorize Loop 4 on the hybrid base — CANCEL the proposed 8–14 h procedural weight-paint

(Does not authorize the remaining loops, human testing, merging, backend, or M1.)

Decisive evidence — identical scene, same motions

Sit	Walk

A (procedural, white) creases at the hips (sit = stiff crouch) and walks segment-y. B (Quaternius, yellow) — a 65-joint pro rig with 43 CC0 clips (Idle/Walk/Sitting/PickUp_Table/Push) — sits and walks with clean natural deformation. A fails the advance rule (hip deformation 2/5, overall 3/5); B/C pass (overall deformation 5/5).

Why hybrid, not B-as-is

Recolor + apron/toque overlays work (), so restaurant identity is feasible — but currently crude, not yet credible (props are offset boxes; needs conformed clothing/hair + bone-attached props). That customization is Loop 4.

Labor finding (the point of the spike)

• Procedural (A) → pass: 15–25 h manual repair, still segment-y.
• Hybrid (C) → pass: ~10–18 h, clean deformation and custom identity.
  → The 8–14 h procedural weight-paint should not be funded; the CC0 foundation is cheaper and better.

License / hygiene

Quaternius UAL = CC0 (re-verified at source: commercial + modification + redistribution). Raw mesh kept git-ignored; only renders committed; the customized derivative gets committed in Loop 4.

Full analysis + scoring table + all renders: foundation-spike/FOUNDATION-DECISION.md on spike/v0-foundation-comparison.

Art PM decision requested: confirm Candidate C (hybrid) as the foundation and cancel the procedural manual pass; Loop 4 then builds the customized character on the UAL rig. Keep #8 open/unmerged; human test deferred.