feat(context-guard): ground the default zone boundaries in research, per model #6227

Problem

context-guard's default zone boundaries are not grounded in recorded research. Today there are three zones (smart, acceptable, dumb) with these defaults: smart up to 50% and acceptable up to 75% by percentage; by tokens, 100K/160K on a 200K window and 200K/400K on a 1M window (plugins/context-guard/hooks/zone.ts:10-17). The owner works to different bands (about 15%, 40%, 60% as the compact-or-handoff target, 80% as the wrap-up limit), and those bands come from community practice (largely Matt Pocock's), not measured data.

Decide, then build

• Research what context fill degrades quality, per current model (Opus 5.5, Fable 5.1, Sonnet 5.5, and Haiku when its next version ships), in this order: Anthropic guidance first (if Anthropic publishes a number, that is the default), then published long-context evaluations, then recognized practitioners, with dates and dissent noted.
• Whether the defaults should vary by model as well as by window size, and whether a fourth zone (a wrap-up band before the hard limit) earns its place.
• Align with the repo's existing compact-versus-handoff guidance (session-flow skills) so the guard's zones and that guidance name the same thresholds.

Acceptance

• Each default boundary in zone.ts cites its basis (URL and date, or judgment).
• README and CHANGELOG record any change, with a version bump.

Related

• fix(guards): band rows and notice repeat the status line, and the documented /<guard>:band commands do not exist #6222 (guard display), docs(conventions): minimal, user-visible model-facing text for hooks and mods, then sweep #6225 (model-facing hook text).