test(bench): add graph.html render benchmark + baseline

[muqsitnawaz]

A reproducible benchmark for the interactive graph.html viewer, so the rendering optimizations coming next can be measured, not asserted. No product code changes — adds a bench/ harness only.

What it measures

Per graph size, without modifying the generated HTML:

• payload — graph.html size on disk
• DOMContentLoaded — document load/parse time
• time-to-settle — wall-clock until the force-directed layout stops animating (detected by wrapping requestAnimationFrame; settled = no frame scheduled for 400 ms)
• peak JS heap — performance.memory.usedJSHeapSize

Baseline (this machine) — the O(N²) main-thread force sim, quantified

nodes	edges	payload (MB)	DOMContentLoaded (ms)	time-to-settle	peak heap (MB)
1,005	2,400	0.81	152	2.3 s	9.5
5,001	12,000	3.89	175	14.9 s	12.1
10,005	24,000	7.8	219	35.2 s	24.8

Layout settle scales ~quadratically (5x nodes → ~6.5x time); a 10k-node graph pins the main thread for 35 s on every page open. This is the number the follow-up server-side-layout PR targets (precompute in C++ → browser renders a static layout, settle → ~0).

Pieces

• bench/gen_tree.py — deterministic synthetic Python project of ~N nodes, grouped into packages (communities) with cross-file calls/imports.
• bench/bench.mjs — Playwright driver; loads a graph.html, detects settle, prints one JSON row.
• bench/run.sh — generates a tree per size, scans with cgraph, runs the driver, writes baseline-report.md.

Requires Python 3, Node, and Playwright chromium (bunx playwright install chromium). Part of a rendering speed-up series (dependency-free; server-side igraph layout + Canvas culling to follow).