fix(skills): make skill loading explicit and repeatable

[danielkov]

Summary

Rename the Agent Skills tool from activate_skill to skill, declare its text output schema, and return full instructions on every read. Bump agentkit-tool-skills to 0.10.8 and update the umbrella dependency and documentation. Replace the timing-sensitive Runlet concurrency probe with deterministic barrier synchronization.

Motivation

Consumers that run the skill tool behind a composition boundary must forward its result for the instructions to reach model context. The previous activation framing and missing output schema encouraged callers to treat the operation as a one-time side effect; session deduplication then made accidentally discarded instructions unavailable on retry.

The existing concurrency test relied on a bounded yield loop, so a slow scheduler could complete one probe before the other entered and report a false failure.

Impact

The advertised tool name changes to skill, so consumers with hard-coded activate_skill lookups must update. Repeated reads now return the complete skill instructions instead of Skill already read.

Technical details

Compatibility

SkillRegistry::reset_activations remains as a deprecated no-op for patch-release source compatibility.

Deterministic concurrency check

The Runlet scheduling test now holds both independent calls at a two-party barrier. A five-second timeout still fails promptly if the calls are actually serialized.

Validation

cargo test -p agentkit-tool-skills, cargo check -p agentkit --features tool-skills, and cargo test -p agentkit-tool-compose --all-features pass. The formerly flaky concurrency test passed 30 consecutive targeted runs, and mdbook build book succeeds.