v0.5.1 — TUI honesty + accurate skill cost

What's fixed

• TUI bindings actually fire on a fresh tab-switch. Every screen focuses its primary DataTable in on_show, so per-tab keys (Sessions d, Perms p/D, Schedule a/r/n/p, etc.) work without a hidden focus dance.
• Sessions d is real — preview + confirm modal + audit-logged soft-delete via core.sessions.delete_session.
• Perms p and D are real — new preview modals for prune-stale and dedup, with bulk remove_rule calls that back up each settings.json before write.
• Schedule key naming is consistent across BINDINGS, header, F1 help and the empty-state CTA: a add, r remove, n run-now, p promote.
• Skill cost is accurate. enabled_skills used to count full SKILL.md file size in tokens; now respects skillListingMaxDescChars (default 1536) and skillOverrides ("off", "name-only") from ~/.claude/settings.json. Heavy skills previously over-reported ~16x.

Dropped (sham bindings)

• Sessions / search → use cc-janitor session find
• Audit u undo → use cc-janitor undo

Tests

319 passed (was 302). +17 new tests covering focus, sessions delete, perms preview modals, and skill cost calculation.

🤖 Generated with Claude Code