skillopt-sleep: make transcript sources and backends pluggable (entry points), and fail on unknown backend names #314

Problem

harvest_sources.harvest_for_config() and backend.get_backend() are fixed if-chains. Every new harness needs a change to core: open right now are #153 (Antigravity), #266 (DeepSeek Harness, from #225), #282 (Codex CLI rollouts), #114 (Hermes as backend and source), plus #97 (VS Code Copilot). Also, get_backend() returns MockBackend() for any name it does not know, so a typo or an out-of-tree backend silently produces a 0.0 -> 0.0 night instead of an error.

Motivating case

On one macOS machine we run a nightly cycle over claude, codex, cursor, opencode and copilot_cli natively, plus three stores upstream cannot read:

• zcode: an OpenCode fork whose db.sqlite lacks the agent and metadata columns harvest_opencode requires, so it finds 0 sessions.
• Antigravity: per-conversation SQLite with protobuf steps; we also read the antigravity-cli and antigravity-acp conversation dirs, which feat(sleep): Antigravity transcript harvest source (re #97) #153 does not.
• Crush: a per-project crush.db listed in projects.json.

Today these reach the engine only by pre-writing digests.json into the handoff dir before round 1. Every model call goes through the handoff backend plus an external answerer, because no backend can be registered from outside. The adapters and the runner are public: https://github.com/djbclark/djbclark-ade/blob/master/bin/skillopt_sources.py and https://github.com/djbclark/djbclark-ade/blob/master/bin/skillopt-night

Proposal

1. Two entry-point groups, consulted after the built-ins: skillopt_sleep.sources (name -> harvest(cfg, *, since_iso, limit, scope, invoked_project) -> list[SessionDigest]) and skillopt_sleep.backends (name -> factory(cfg) -> Backend). --source and --backend accept any registered name, and choices= lists come from the registry.
2. --source-module pkg.mod:func and --backend-module pkg.mod:Class for unpackaged adapters.
3. get_backend() raises on an unknown name instead of returning MockBackend (keep mock explicit).
4. Key the harvest checkpoint by (project, source) (skillopt-sleep: harvest checkpoint is keyed by project, so a second --source in the same project skips its unseen sessions #312) so several sources can feed one project.
5. A small public helper API (is_headless_replay, is_agent_session, is_meta_prompt, detect_feedback, redact_secrets), so external sources apply the same filters and redaction as built-in ones.

The existing sources would register through the same mechanism, and per-source PRs such as #153 and #266 would become optional plugins instead of core changes.

Offer

I am happy to send a PR implementing 1-3 and 5 with tests (keeping every current --source and --backend name working), and 4 separately if #312 is accepted. Tell me if you prefer a different shape, for example a config-file registry instead of entry points.

Version context: main at 343db22.