perf: cache contextWindow in metrics transform instead of per-emit prefix scan

[kianwoon]

Performance — Tier 2 (Medium Impact)

File: src/server.ts:346

Problem

getContextWindow() is called on every throttled WS emit (~4 calls/sec). It does a linear prefix scan on the model name to determine the context window size. The model does not change mid-stream, so this is wasted work.

Fix

Cache contextWindow once when actualModel is resolved (start of transform), store on the transform instance, and reuse for all subsequent emits.

Impact

Eliminates ~4 linear scans/sec during active streaming.

Source: Performance review (2026-04-25)