/foundry:learn-distill

The self-improvement consumer — a native-scheduling WRAP for the cadence. Two halves:

• CUSTOM (no native equivalent): the deterministic clustering + report logic, implemented in scripts/foundry-distill.py — read the dated partitions under .foundry/session-learnings/<YYYY-MM-DD>/*.jsonl (the reconciliation .harvest-log.jsonl + dotfiles excluded), token-overlap cluster (≥3 shared normalized tokens drawn from every string scalar field + tags; no embeddings; fully-pinned canonicalization + ordering → byte-reproducible), threshold-promote (≥3-record components) to HBK/memory/skill/unclassified candidates over the real type/kind vocabulary, write .foundry/learnings/DISTILL_REPORT-<UTC-date>.md (an untracked runtime artifact — NOT the governed specs/ tree), then prune buffer partitions whose directory-NAME date is >90 days old (never by mtime). Run it directly: scripts/foundry-distill.py [--date YYYY-MM-DD] [--dry-run]; self-test scripts/foundry-distill.py --selftest. (The doctor drop-in check this line originally named was retired with the drop-in registry in the v0.25.0 realignment — an open loop is now visible only by running the distiller itself.)
• WRAP over native: the CADENCE. Instead of a bespoke scheduler, the distill tick is scheduled via native ScheduleWakeup (session-local) or CronCreate (persistent) — e.g. a daily tick that runs the consumer. Producers feed the buffer via /foundry:learn-capture (the in-session path) + the worker-sidecar harvest.

Procedure

1. Scheduled tick (native CronCreate daily, or ScheduleWakeup): invoke the consumer — scripts/foundry-distill.py.
2. Consume the buffer (Read → Cluster → Threshold → Write → Retention, in that strict order — retention prunes only partitions whose NAME-date is >90 days old, never a within-window partition).
3. Emit .foundry/learnings/DISTILL_REPORT-YYYY-MM-DD.md with HBK/memory/skill/unclassified candidates (operator-reviewed; no auto-dispatch).

Anti-patterns

• A bespoke scheduler — use native ScheduleWakeup/CronCreate for the cadence.
• Embeddings / semantic clustering — clustering is deterministic token-overlap (reproducible across runs).
• Running retention before the read phase — order is non-negotiable.
• Auto-dispatching candidates — the report is a recommendation; the operator drives follow-ups.