Feat/canonical model and tool formats

[Nivesh353]

Summary

Standardizes two declarative formats in the gitagent spec so exported agents
work correctly on downstream runtimes (notably gitagent, which silently
dropped tools/models in the old formats):

1. Model field is now canonical provider:model (e.g. anthropic:claude-sonnet-4-6)
2. Tool implementation now uses script: (with path: kept as a deprecated alias)

1. Canonical provider:model model format

model.preferred / model.fallback must now be provider:model — bare names
(gpt-4o) are rejected. A shared parseModel() util (splits on the first colon)
replaces the per-adapter provider-guessing logic.

• New: src/utils/model.ts — parseModel(), KNOWN_PROVIDERS, isKnownProvider()
• Schema: agent-yaml.schema.json enforces ^[a-z0-9-]+:.+$ on preferred/fallback
• Validation: errors on bad format, warns on unknown provider
• Adapters/runners reformat to each target's needs (bare for claude/openai/kiro;
  provider/model for opencode/github/nanobot/openclaw; {id, provider} for gemini;
  passthrough for gitclaw). Deleted the old inferProvider/mapModelName guessers.
• import.ts now emits canonical provider:model
• Migrated all example agent.yaml files + spec/docs examples

2. Tool implementation.script

type: script tools now use script: (the form gitagent's runtime reads).
path: is still accepted but deprecated.

• Schema: tool.schema.json adds script, deprecates path, requires one-of + runtime
• Migrated example tools + spec doc

Breaking change

Agents with bare model names (gpt-4o) or no provider prefix will now fail
validation. Migrate to provider:model (e.g. openai:gpt-4o). Tool path:
remains valid but should migrate to script: for gitagent compatibility.