Skill sources, generated assets, and PR synchronization

Canonical skills live under skills/; .agents/skills/ is generated by scripts/sync-skills.mjs, so edit canonical files and regenerate. OpenCode memoizes discovery and does not hot-reload SKILL.md; verify with opencode debug skill in a fresh process. In clean worktrees, generate prerequisites in repository order: generate:schema, then generate:docs, then generate:sdk; missing generated files are not product regressions. On getsentry/cli PR pushes, docs regeneration may add a bot commit and force-advance the branch, so re-fetch, fast-forward, and validate the remote head rather than reverting generated output. DEVELOPMENT.md prose has no automated staleness check.