feat: custom pipeline stages + pluggable handlers

[terraboops]

Summary

Ships the custom-stages spec in full: user-defined stages → role_groups → roles pipelines, JSON-in/out handler contract (agent / script / webhook / human / k8s_job), atomic role execution via baseline+rollback, watcher scope with done-stage reactivation, and the UI surfaces to operate it all.

• New data model — stages[] + role_groups[][] with done sentinel, backwards-compatible with today's agents/post_ready/parallel_groups via on-read migration.
• Handlers — Handler ABC with five concrete types; scheduler dispatches through next_roles_in_current_stage (role-group parallelism) and mirrors state into a role_state tree.
• Retry + timeout — wraps any handler with the spec's exponential backoff ladder (30→60→120→240→300 cap) and terminal failure state.
• Atomicity — BlackboardSnapshot + crash recovery on startup, git used as a patch primitive (no .git dirs in idea folders).
• Watchers — positive-list scope ("*" / role-names / done) and reactivate + add_stage payloads for re-opening done ideas.
• UI — stage badge, nested progress bar, handler-type badges, failed-role retry panel, + Add stage affordance + rename endpoint.
• API + CLI — POST /ideas/{id}/roles/{role}/retry, POST /ideas/{id}/stages, POST /ideas/{id}/stages/{name}/rename, trellis retry CLI.

See FEATURES.md for the exhaustive 176-behaviour catalogue.

Test plan

• 453 pytest cases across 9 new test files (unit + integration + HTTP + CLI)
• 9 Playwright flows against the live server in tests/test_browser_custom_stages.py
• Full suite green locally; ruff + ruff-format clean
• Manual smoke test on my-brain with 3 seeded ideas (multi-stage, failed-role, done-stage) — screenshots captured and reviewed
• Merge to integration + run alongside PR feat: SPIFFE/SPIRE identity federation for Git forge access #1 (SPIFFE) and PR Add Prose format support for pipeline templates #2 (Prose)

🤖 Generated with Claude Code