feat: promotion candidate detector and pk candidates list|accept|reject

[GQAdonis]

Summary

Change C1a of team-aware-learning-memory-impl. Promotion beyond a project is now auto-proposed and human-confirmed (skill-system design §3).

• Worker detector: pk-learning-worker/src/promotion.rs runs after the job loop in run-once.
  • It fingerprints lessons as word-trigram shingles into ~/.prometheus/learning-index/lessons.jsonl, with project and source evidence.
  • It proposes a candidate when a lesson recurs (Jaccard ≥ 0.6) across ≥ 2 projects, is tagged global without a [GLOBAL]/[USER] marker, or names a dependency or CLI and has no repo-relative path.
  • Candidates are written atomically to ~/.prometheus/promotion-candidates/pending/<id>.json, with an evidence list. Proposal is idempotent, and a candidate already accepted or rejected is never re-proposed.
• New subcommand pk candidates list|accept|reject --kind promotion|skill (frozen surface: additive):
  • accept upserts the lesson into the shared KB as promoted-<id> and commits the shared prompt snapshot.
  • It then queues a schemaVersion 2 add_memory op in the learning-queue outbox, keyed @global/@global (or @user:<hash>/@user for a user-scoped candidate). payloadHash is the SHA-256 of the arguments.
  • Finally it moves the file to accepted/. reject moves the file to rejected/.
  • For --kind skill, list and reject work. accept exits non-zero with "not yet supported".
• OpenSpec: openspec/changes/promotion-candidates/. It covers the proposal with its consumers table, the design (including a known race and the noise trade-off), the spec, and the tasks.
• No dependency change. Cargo.lock is unchanged.

Commits

• ec23551 RED: the OpenSpec change and both integration tests. They were not executed at that commit, because the repository's implementation-first policy defers test runs.
• d7905e4 GREEN: the implementation, plus rustfmt applied to the test file.

Test plan (local, serialized Cargo)

• verify.sh runs cargo test -p pk-learning-worker --test promotion (1 passed) and cargo test -p pk-cli --test candidates (2 passed): verify OK.
• cargo test --locked -p pk-cli --test context --test context_scoring --test tags_and_type: 3 + 3 + 1 passed.
• cargo test --locked -p pk-learning-worker --test attribution --test worker: 2 + 5 passed.
• cargo clippy --locked -p pk-learning-worker --bins --test promotion --test attribution --test worker -- -D warnings: clean.
• cargo clippy --locked -p pk-cli --all-targets -- -D warnings -A clippy::too_many_arguments: clean.
• Verifier gate: cargo fmt --all -- --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked. Two failures already exist on origin/main, and the gate will hit them:
  • pk-cli/src/main.rs::run_context trips clippy::too_many_arguments (8 of 7).
  • pk-learning-worker's bin unit tests do not compile, because the learning_job helper near main.rs:1711 lacks project_id, team_id and role_id after fix(worker): attribute learning to project/team/role and commit the prompt snapshot #32.
• Downstream handoff note for prometheus-skill-system: the pk pin, docs/guide/13-tools-reference.md, and kbd-open.sh/reflector surfacing of pending candidates.

Review status: builder-context only, not independent.

🤖 Generated with Claude Code