viz test: plotly integration

Hidden test page — real experiment data across all runs. Drag on any axis to filter trajectories.

---

Cross-Experiment Comparisons

Parallel Coordinates: Old vs New Sonnet (swapped)

Notebook and Viz

Width Bar: All Strategies (sonnet 4.6, multi-fetch)

Radar: Behavioral Profile (old vs new sonnet)

2D Scatter: Tool Calls vs Cost

3D Scatter: Tool Calls × Cost × Duration

Strip Plot: Per-Dimension Distribution

---

Per-Experiment Parallel Coordinates

Part 1 — Naive Environment

none / old-sonnet (control, K=2)

naive / old-sonnet (K=10)

swapped / old-sonnet (K=10)

swapped / sonnet 4.6 (K=10)

---

Part 2 — Multi-Fetch Environment (sonnet 4.6)

progressive (K=5)

rescue (K=5)

breadcrumb (K=5)

breadcrumb_poison (K=5)

breadcrumb_exec (K=5)

url_redirect (K=5)

url_redirect_funky (K=5)

ipa_progressive (K=5)

ipa_exfil (K=5)

base64_breadcrumb (K=5)