Scope platform support honestly and centralize model IDs behind capability tiers

[Copilot]

The README led with "works across Claude Code, GitHub Copilot," but model routing, subagent dispatch, and required tools were Anthropic/Claude-specific — on Copilot the triage → dispatch → verify chain degraded silently to "one agent does everything." Dated model IDs (e.g. claude-opus-4-7) were also duplicated across the setup routing tables, a runtime-failure and maintenance risk.

Changes

• Honest platform claim (README): Reworded the tagline — Claude Code is first-class, GitHub Copilot is explicitly experimental. Added a Platform compatibility table mapping each Claude primitive to its Copilot/generic equivalent so behavior degrades predictably, not silently.

  Ritus concept	Claude Code	GitHub Copilot / generic
  Skill invocation	Skill tool	/command or read SKILL.md
  Subagent dispatch	fresh subagent (programmatic)	separate pass in clean context
  Model routing	per-tier model at dispatch	advisory — pick model in client

• Single model-ID surface: The three routing tables in skills/setup/SKILL.md now reference capability tiers (haiku/sonnet/opus) instead of dated IDs. Concrete IDs live once in a model resolution table (tier → ID per platform, with a real GitHub Copilot column) and in docs/profiles/runtime.yml (model_ids), rendered into PROJECT_CONTEXT.md. IDs are labeled examples with a "set to your platform's current model" instruction to prevent rot.

• Abstracted Claude-only tool names: start-ritus (Skill tool, subagent dispatch) and pr-review (WebFetch) now use platform-neutral wording with the Claude-specific name noted; remaining names (AskUserQuestion, CLAUDE_PLUGIN_ROOT) already had fallbacks and are documented in the compatibility table.

Notes

• runtime.yml schema_version bumped 2 → 3 for the new model_ids field.
• Tier tokens (haiku/sonnet/opus) are retained across skills and reframed as capability tiers resolved via the profile, rather than renamed — keeps the change surgical while removing the dated IDs.
• On Copilot, model_routing is documented as advisory since there is no programmatic dispatch; verify/review run as separate clean-context passes.