feat: proactive enrichment-recall subsystem (the "🧠 Help me remember" loop)

[jordanrburger]

Implements the design spec merged in #150 (docs/superpowers/specs/2026-06-21-enrichment-recall-subsystem-design.md) — the pull-capture half of the feedback loop. At the end of a dreaming run the system appends a small ranked set of "🧠 Help me remember …" questions to its wrap DM, asking about facts no connector can see (in-person events, decisions, head-knowledge), and later feeds the answers back into the KB.

What's here (matches the spec's 4-part plan)

1. Generator → templates/scripts/generate-enrichment-questions.py, wired into _CAT1_FILES_FROM_PLUGIN exactly like recurring-task-status.py (raw .py, copied verbatim, no template rendering). A pure-stdlib, read-only KB scanner that ranks connector-blind candidates (rank 0 explicit [needs:] flags → rank 4 thin stubs), with one-run --exclude rotation and a persistent --reject-file stoplist.
2. State seeds (install-only, never clobbered on upgrade) → scripts/enrichment-stoplist.txt + knowledge-base/enrichment-qa-log.md, added to _INSTALL_ONLY_TEMPLATES. Shipped as empty headers.
3. Recall section → phases/modes/feedback-processing.md (dreaming-scoped): generate → quality-gate → compose the "🧠 Help me remember" block → record dismissals to the stoplist and rounds to the Q&A log. New Step 1f; old commit step renumbered to 1g.
4. Tests → engine/tests/unit/test_enrichment_questions.py.

De-personalization (this is the public engine)

The generator was ported from a running instance and scrubbed: stripped dated user-feedback quotes, real names, employer references, and Pattern #NN citations; genericized the argparse/header string literals.

⚠️ Please scrutinize the regex generalization (the spec called this out as the subtlest change): the connector-answerable guard RESEARCH_ANSWERABLE previously hardcoded the employer name in an alternation (…|<employer> customer|…) plus product-specific jargon. It's now employer-agnostic (customer status, account id, crm record, deal stage, renewal date, attribution phrasing, named-non-user-person phrasing, connector/profile identity facts) — aiming to keep the connector-answerable-vs-head-fact discrimination without baking in any org/product vocabulary. A scrub check confirms no leaked identifiers.

Verification

• 883 passed (full engine unit suite) — incl. bootstrap install/upgrade wiring.
• ruff check clean on the new/changed files.
• New tests cover: ranking/dedupe, persistent stoplist, --exclude rotation, the guards (research-answerable dropped, projects/ + ontology/entities/ scoped out, resolved gaps skipped), the read-only contract (asserts the KB is byte-identical after a run), bootstrap wiring, and assembly scoping (recall lands in DREAMING, not SKILL/RESEARCH).

Open questions carried from the spec

• The engine has no dedicated dreaming notification-composition phase; the recall attaches to feedback-processing.md (the closest existing home). A future consolidation would be a cleaner home — out of scope here.
• enrichment-qa-log.md growth is unbounded today; a retention policy is a possible follow-up.

🤖 Generated with Claude Code