feat(skills): execute session-scoped command hooks

[nonoqing]

Summary

Claude-format skills with command hooks were rejected as invalid during discovery. Invoking a supported skill now activates its command hooks in the existing session Hook engine, and the UI distinguishes unsupported capabilities from actual scan failures.

Type and Areas

Type: Feature / bug fix

Areas: Rust Agent Runtime, Core skill/tool/permission/session owners, Desktop skill DTO, Web UI diagnostics, locales, docs, Cargo boundary contracts.

Motivation / Impact

• Register hooks only on skill invocation, isolate them by session/workspace, preserve idempotence and once, and cancel/remove them during session teardown.
• Match Claude Bash hooks to ExecCommand, translate Bash/Write inputs and updatedInput, and retain original-input constraints and permission denials. Hook ask requires a fresh reply through the existing permission mailbox, including in bypass mode.
• Make skill activation a preflight barrier so later tools in the same model response see the new hooks. Unknown execution semantics reject the entire skill.
• Keep discovery/import passive and retain imported Claude dialects. Split unsupported-capability notices from missing-file/format errors in both skill surfaces.

Verification

270 focused Rust tests passed (Core pipeline 52, Core hooks 19, Core skills/imports 82, interaction contracts 47, process contracts 29, skill contracts 41):

cargo test --locked -p openbitfun-core --no-default-features --features agent-runtime,git,external-sources --lib agentic::tools::pipeline::tool_pipeline::tests
cargo test --locked -p openbitfun-core --no-default-features --features agent-runtime,git,external-sources --lib native_hooks
cargo test --locked -p openbitfun-core --no-default-features --features agent-runtime,git,external-sources --lib agentic::tools::implementations::skill
cargo test --locked -p openbitfun-agent-runtime --no-default-features --features agent-runtime --test native_hook_execution_contracts --test agent_interaction_contracts
cargo test --locked -p openbitfun-agent-runtime --no-default-features --features agent-runtime --test agent_definition_contracts skill
cargo check --locked -p openbitfun-agent-runtime --no-default-features --features definition-contracts

After integrating current upstream/main, these also passed:

pnpm --dir src/web-ui run test:run src/app/scenes/skills/hooks/useInstalledSkills.test.tsx src/flow_chat/components/ChatContextPickerOverlay.test.tsx src/infrastructure/api/service-api/ConfigAPI.test.ts
pnpm run check:web
pnpm run i18n:audit
pnpm run check:core-boundaries
node scripts/check-git-object-sizes.mjs --base upstream/main --head HEAD

The three frontend files passed 43 tests; i18n reported zero warnings. The architecture checker regression suite also passed 148 tests.

Reviewer Notes

AI-assisted implementation with focused automated testing. No packaged-app UI capture or full multi-device E2E was performed. Local execution and remote-workspace rejection are covered; Remote Control, Peer Device, and Detached Dispatch E2E remain unverified.

Only synchronous command handlers and the engine's existing lifecycle events are supported. Project skills require the project-hook gate. Registrations are process-local and must be reactivated after a runtime restart. SSH/remote-workspace activation fails explicitly without running commands on the controller. Native hooks.json retains its existing Codex decision contract.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.