feat(skills): support explicit invocation arguments

[limityan]

Summary

• honor Claude-compatible user-invocable and argument-hint metadata in Web and CLI/TUI invocation surfaces while keeping user and model visibility independent
• accept optional Skill tool arguments and expand the loaded prompt through one shared static helper reused by the existing Claude command adapter
• align Web and CLI/TUI insertion on [$skill-name], showing argument hints without inserting them into user input

Boundaries

• keeps direct Skill discovery and existing registry precedence/load paths; no import records, source graph, coordinator, cache, watcher, or new settings
• does not add Plugin Runtime, hooks, MCP, customization, dynamic shell execution, named arguments, or allowed-tools/fork/agent/model semantics
• leaves management, mode enablement, permission identity, and local/remote workspace ownership unchanged

Review fixes

• complete the internal SkillInfo test fixture so the full agent-runtime test target compiles
• preserve backslashes before non-placeholder dollar expressions such as \$HOME, while retaining the documented single- and double-backslash behavior for supported placeholders
• add owner regression coverage for ordinary variables, invalid indexed placeholders, and dynamic-variable-shaped text

Validation

• cargo check --locked --workspace
• cargo test --locked -p bitfun-agent-runtime
• cargo test --locked -p bitfun-services-core --features markdown --test markdown_owner_contracts
• cargo test --locked -p bitfun-claude-code-adapter --test command_source
• cargo test --locked -p bitfun-core skill_ --lib -- --nocapture
• cargo test --locked -p bitfun-cli skill_selector -- --nocapture
• cargo test --locked -p bitfun-cli -p bitfun-acp -p bitfun-agent-runtime --no-run
• pnpm run type-check:web
• pnpm --dir src/web-ui run test:run src/flow_chat/utils/skillPromptReference.test.ts
• node scripts/check-core-boundaries.mjs
• pnpm run check:repo-hygiene
• git diff --check

A fresh independent adversarial review passed with no P0-P2 findings after the two review issues were fixed. The full Windows CLI execution additionally reached five dispatch file-lock/concurrency failures in untouched code; those failures are not counted as successful validation, while the exact combined compile step above passes.