GLASS CATHEDRAL — FRACTURE DYNAMICS BENCHMARK v1.0

Mühendislik değerlendirmesi kodun sözleşmeye uyumunu, solver ve temas fiziğini, faz nedenselliğini, metrikleri ve görsel açıklığı birlikte ele alır. Uygulama özpuanından ve katı prompt geçerliliğinden ayrıdır; maliyet ve üretim süresi puana dahil değildir.

# GLASS CATHEDRAL — FRACTURE DYNAMICS BENCHMARK v1.0
### SINGLE-FILE · SELF-GRADING · 75 SECONDS

Build **ONE self-contained HTML file** simulating a Gothic stained-glass cathedral facade that fractures under impacts, wind, and support failure. Run **5 automated phases × 15 seconds = 75 seconds** and grade the results out of 100. No user interaction.

## 0. HARD CONSTRAINTS
- **Vanilla JavaScript + Canvas 2D**, all inline; no libraries, CDNs, external assets, build steps, or network. Must run offline by double-clicking the HTML.
- **Fixed physics dt = 1/120 s**, accumulator, simulation-time phases. Rendering and canvas size must not change physics.
- Seeded PRNG (`SEED = 1337`) for all randomness; **no `Math.random()`**. Same-seed runs must reproduce physics metrics and end-state hash.
- **No scripted cracks, fake shattering, teleported fragments, forced trajectories, or hard-coded PASS values.** Damage must result from forces and bond failure.

## 1. SCENE & PHYSICS
- Draw **three tall pointed-arch stained-glass windows** inside a dark stone cathedral, with luminous red/blue/gold panes and a solid floor. Use a fixed 2D world coordinate system.
- Use **150–200 irregular triangular glass shards** total, initially connected as solid panels. Each triangle is a movable rigid body with position, orientation, velocity, mass, and inertia. This is a simplified **bonded-shard fracture model**, not full finite-element glass mechanics.
- Neighboring shard edges are joined by **two breakable XPBD anchor constraints** (resisting translation and rotation). Implement real compliance, accumulated multipliers, and positional/angular corrections: `alphaTilde = compliance / dt²`, `deltaLambda = (-C - alphaTilde * lambda) / (effectiveInverseMass + alphaTilde)`.
- Bonds **irreversibly break** above measured tensile/shear load or strain thresholds, with ±10% seeded strength variation. Broken bonds must create real separation. Define a *detached shard* as one no longer connected by intact bonds to a frame support.
- Simulate gravity, damping, **physical sphere–glass impact impulses**, wind acting on exposed area, and collisions with floor, frame, and other shards (spatial-grid broadphase). Reconstruct velocities after XPBD corrections; use 4–8 solver iterations per step.
- Never silently remove glass, restore damage, or hide NaNs/penetration. Glass mass and triangle count are conserved.

## 2. AUTOMATIC 75-SECOND CHOREOGRAPHY
| Phase | Time | Event | Measurements |
|---|---|---|---|
| **P1 — STILLNESS** | 0–15 s | Intact glass settles under gravity. | Drift, intact bonds, stability |
| **P2 — FIRST IMPACT** | 15–30 s | A steel sphere physically strikes the **left window**. | Contact impulse, local crack growth |
| **P3 — SECOND STRIKE** | 30–45 s | A faster sphere strikes the **right window**; loose pieces hit the floor. | New breaks, detached shards, collisions |
| **P4 — CROSSWIND** | 45–60 s | Wind blows right, reverses left, then becomes gusty. | Signed lateral motion, stress, FPS |
| **P5 — SUPPORT FAILURE** | 60–75 s | Release **only the predefined upper support group** of the center window. Its glass must fall, break, and collide through simulated physics. | Cascade, fallen shards, conservation |

The director may launch spheres and trigger the **single explicit support release**. It must NOT directly create cracks, delete triangles, or animate falling glass.

## 3. VISUALS & HUD
- Cinematic fixed camera; colored translucent glass, dynamic **stress heatmap** (cool → yellow → red), gaps along truly broken bonds, rotating/falling shards, visible projectile paths. Decorative dust is optional and never graded.
- Always show **phase/countdown, 75 s timeline, FPS, broken bonds, detached/fallen shards, impact impulses, maximum stress ratio, and live score**.

## 4. SELF-GRADING — 100 POINTS
Grade from **measured time-series**, not visual appearance. Split each phase's points equally across its checks; show individual PASS/FAIL values and partial credit.

| Phase | Points | Checks |
|---|---:|---|
| **P1** | **15** | No NaN/Inf; shard count preserved; <1% broken bonds before impact; average final-3-second drift <0.03 world units |
| **P2** | **20** | Positive measured contact impulse; ≥5 new bonds broken near impact; ≥60% of other windows' bonds intact |
| **P3** | **20** | ≥5 additional broken bonds; ≥5 shards detached from frame; actual floor contacts without tunneling |
| **P4** | **20** | Mean lateral glass response reverses sign with wind; no invalid states; average rendering FPS ≥30 (also report FPS separately) |
| **P5** | **25** | Correct support release recorded; ≥20 center-window shards fall below their original sill level or touch ground; glass mass/count remains 100%; no invalid states |
| **TOTAL** | **100** | |

**Hard fail = 0/100** for fake fracture animations, absent XPBD bond solver, missing phases, or fabricated metrics. All initial parameters and thresholds must remain fixed for the entire run and be included in the output.

## 5. FINISH & OUTPUT
- At **t = 75 s**, freeze physics and show an on-screen table: measured values, thresholds, PASS/FAIL, points, **TOTAL /100**.
- Print exactly one final JSON line: `console.log('BENCHMARK_JSON:' + JSON.stringify(scorecard))` containing `seed`, `phaseMetrics`, `brokenBonds`, `detachedShards`, `groundContacts`, `impactImpulses`, `massPercent`, `invalidStates`, `avgFPS`, `stateHash`, `totalScore`.
- Calculate `stateHash` from quantized final positions, orientations, velocities, and bond states. Include **Restart** and **Pause** controls.
- Keep code concise and functional: PRNG → glass bodies → XPBD bonds → collisions → director → metrics → rendering → grading.

**Deliver ONLY the complete HTML file.**

*Physics basis: [XPBD, Macklin et al. (2016)](https://doi.org/10.1145/2994258.2994272). The fracture model is a practical benchmark approximation, not a validated predictive glass-engineering solver.*