Add dynamic skill loading: batch ops + capacity reporting (#52)

[devrimcavusoglu]

Summary

Phase 1 of #52 — dynamic skill loading. Lets agents manage their working set of installed skills against per-platform capacity thresholds.

• Batch install/uninstall: skern skill install a b c --platform <p> accepts multiple skill names; per-skill outcomes in skills[]. Partial failures don't abort the batch; non-zero exit only when every op fails.
• Capacity block in every install/uninstall JSON response: {platform, scope, installed, threshold, headroom, over_budget} — agents react without an extra query.
• --enforce-budget flag on install: refuses the operation when the resulting count would exceed the per-scope threshold.
• skern skill list --with-platforms: enriches each entry with installed_on[] (detected platforms where the skill is currently installed at the registry skill's scope).
• --platform all removed (D6): each invocation targets exactly one platform; agents specify the platform they're running on. --platform all now exits with a validation error.
• Per-platform per-scope thresholds lifted into internal/skill/capacity.go (project=20, user=50; project ≤ user invariant tested).
• Smoke tests fixed on Windows: os.UserHomeDir() reads %USERPROFILE% on Windows, ignoring bash's $HOME. setup_env now mirrors HOME into USERPROFILE (via cygpath -w) under msys/cygwin/win32. All 57 smoke tests pass.

Spawned #75 (skill stats) as the D5 follow-up.

Decision summary (from issue thread)

ID	Decision
D1	Capacity is advisory; --enforce-budget is opt-in for hard refusal.
D2	LRU state at ~/.skern/state/usage.json — Phase 3, gated.
D3	Batch is the primitive; #48 sync sits on top.
D4	Per-platform AND per-scope (project vs user counted separately).
D5	Skill count is the budget metric. Stats follow-up tracked in #75.
D6	One platform per invocation; --platform all removed.

Breaking change

JSON envelope for skill install / skill uninstall changed shape:

	Before	After
Top level	{skill, scope, platforms[]}	{platform, scope, skills[], capacity}
Per-entry	{platform, success, error}	{skill, success, error}

Bumps next release to v0.2.0.

Test plan

• make fmt clean
• make lint — 0 issues
• make test — all packages green; 8 new tests added (batch happy path, batch partial failure, batch all-fail, enforce-budget, with-platforms, without-with-platforms, batch uninstall + capacity, --platform all rejection)
• make test-smoke — 57/57 passing on Windows after USERPROFILE fix
• Capacity invariant tested (project ≤ user threshold)
• Docs updated: AGENTS.md, commands.md, quick-start.md, agent-setup.md, platform-adapters.md, platforms/index.md

🤖 Generated with Claude Code