feat(distill): distill environment-level memories into a shared catalog

[rockfordlhotka]

Summary

Adds a distill capability to claude-utils: it lifts transferable lessons (shell/OS quirks, CLI gotchas, toolchain, user identity) out of a single project's Claude Code memories and makes them reusable across projects — so Claude doesn't have to re-learn the same environment lessons in every repo.

The design splits judgment from mechanics:

/distill  (skill, Claude)         claude-memsync distill (Go)     /distill-apply (skill, Claude)
  classify + generalize     ──►   rebuild DISTILLED.md index ──►    copy entries into a project's
  write catalog entries           prune stale, report worklist      memory/ + MEMORY.md
  tag originals scope:env               │
        └────────── ~/.claudesync/distilled/  (synced like everything else) ──────┘

What's included

• internal/distill/ (mechanical, Go): BuildIndex regenerates the DISTILLED.md catalog index from per-lesson <slug>.md entry files; Reconcile (--prune) drops entries whose source memory lost the marker or vanished; Preview backs --dry-run; analyzeSources surfaces the worklist of marked-but-not-yet-distilled memories and cross-project conflicts. Includes a tolerant frontmatter parser (handles both the flat and the nested-metadata schemas) with no new dependency. Conservative by design: never prunes when the projects tree is invisible; ignores MEMORY.md and the *.tmp.* litter.
• claude-memsync distill CLI: --prune and --dry-run; loads config with a pre-init defaults fallback.
• Daemon hook: rebuilds the index locally after every sync (startup, debounced, and pull flushes). DISTILLED.md is a derived artifact — git-ignored and regenerated per-PC — so the generated table never causes merge conflicts, while the entry files sync via the existing git add -A.
• Skills (skills/distill, skills/distill-apply): /distill is the classifier of record (classify → generalize → write entries → tag originals scope: environment); /distill-apply seeds chosen entries into the current project's memory.
• init + README: init prints a one-time permission allow-rule for ~/.claudesync/distilled/ (it does not edit the user's global settings.json); README documents the flow, skill install, and permissions.

Design notes

• Classifier of record is the skill, not a frontmatter convention. Claude Code's built-in memory-writer follows a fixed schema and won't emit a scope: marker on its own; rather than depend on that, /distill makes and records the classification, and the marker is the cached result of its judgment. The Go daemon only mechanically aggregates/indexes what the skill produced.
• DISTILLED.md is derived, not synced. Entry files are the synced source of truth; the index is regenerated locally to avoid merge conflicts on a generated table.

Testing

• go build ./..., go vet ./... clean.
• internal/distill has 7 tests (frontmatter parsing both schemas, sorted index, catalog + source conflicts, pending worklist, pruning, never-prune-blind guard) — all passing.
• End-to-end smoke-tested the claude-memsync distill binary against a temp catalog: indexes entries, detects pending, and generates a clean DISTILLED.md.

Follow-ups (not in this PR)

• Optionally let init write the permission allow-rule automatically (behind an opt-in flag, with a careful JSON merge) instead of only printing it.
• The *.tmp.* files in memory dirs look like leftover merge-driver temp files — worth a separate cleanup fix in the sync path.

🤖 Generated with Claude Code