Skill stats: track usage/invocation telemetry for context optimization

[devrimcavusoglu]

Context

Spun off from #52 (decision D5). #52 settled on skill count as the budget metric for dynamic loading because it's already in place and model-agnostic. But for richer context optimization, skern would benefit from per-skill statistics beyond count — e.g. byte size, frequency of installation across projects, time-on-platform, etc.

Idea

Add a skern skill stats command (or extend skern skill show) that surfaces per-skill metrics useful for context-budget decisions:

• Static: SKILL.md byte size, companion file count, total directory size
• Aggregate: how many platforms it's currently installed on, across what scopes
• Historical (if/when Dynamic skill loading: install/uninstall skills on demand for efficient context use #52 Phase 3 lands): install count over time, last-installed timestamp, last-touched timestamp

These stats let agents (and humans) reason about which skills are heavy vs lightweight, frequently used vs stale, etc. — informing eviction and budget decisions beyond raw count.

Open questions

1. Scope of v1: ship just static metrics (size + platform presence) and defer historical data until Dynamic skill loading: install/uninstall skills on demand for efficient context use #52 Phase 3 has a state file?
2. Format: separate skern skill stats command, or fold into skern skill show --stats?
3. Aggregate output: skern skill stats with no args — give a top-N-by-size or top-N-by-platforms-installed table?
4. Token estimation: out of scope per Dynamic skill loading: install/uninstall skills on demand for efficient context use #52 D5, but worth revisiting if a model-agnostic estimator exists.

References

• Dynamic skill loading: install/uninstall skills on demand for efficient context use #52 (D5) — capacity metric decision; this issue is the follow-up for richer stats.