feat(core): dynamic tool output truncation based on context pressure

[netbrah]

What

Make tool output truncation thresholds context-aware. Bump defaults from 25K → 80K chars / 1000 → 2000 lines, with dynamic scaling as context fills up.

Why

The old 25K/1000 defaults were too aggressive early in a session — cutting off useful tool output that the model needs to work with. But keeping a flat 80K budget would be reckless late in a session when context pressure is high.

The fix: scale dynamically based on how full the context window is. Early session → full budget. Late session → shrink proportionally. The model always gets the maximum data it can safely handle at each point in the conversation.

Scaling formulas

• Char limit: min(4 × remainingTokens, baseThreshold)
• Line limit: max(500, floor(baseLines × (1 − usageRatio)))

Both read from uiTelemetryService.getLastPromptTokenCount() (already populated by the existing telemetry pipeline) and the model's context window size.

Changes

File	What
config.ts (core)	Bump defaults, dynamic scaling in getters
config.test.ts	Updated default expectation (25K → 80K)
settings.schema.json	Updated default values
chatCompressionService.ts	Clarifying comment on model-aware window

Test Plan

Existing config.test.ts updated — 79/79 passing. Dynamic scaling verified via getter behavior with uiTelemetryService already wired.

Demo

N/A — internal optimization. Observable via reduced late-session "context too long" errors and richer tool output in early sessions.

Fixes #2566