v0.2.41 — Muse learns you again, and answers faster

Muse learns you again, answers faster, and builds 9x quicker. The
headline is uncomfortable and worth saying plainly: the learning loop had been
mostly inert, and the reason your replies were slow was never Muse.

Fixed

• Your corrections were mostly being thrown away. When you correct Muse, it
  distills a strategy and is supposed to reinforce or retire it as you react. But
  the machinery that decides which rule your feedback is about was calibrated on
  a wrong assumption — it credited 3 of 13 real corrections, and the "you were
  wrong, stop doing that" path had never once fired since it shipped. It now
  credits 14 of 14, with zero mis-attributions, and a correction can finally
  retire a rule that contradicts it. Muse asks its own model which rule you meant,
  and answers "none" rather than guess.
• A rule you wrote yourself can no longer be silently unlearned. Once the
  retire path actually worked, one offhand remark could have wiped a preference
  you typed by hand ("never send an email without showing me the text first").
  Hand-written and evidence-grounded rules are now off-limits to any unattended
  process — only what Muse itself inferred can be retired automatically.
• muse playbook pause now actually pauses. It promised "Muse won't learn
  anything new", but the path that runs at the end of every session ignored it and
  kept writing.
• A fabricated phone number, email, handle or birthday can no longer reach your
  contacts. An invented number "vouched for" by unrelated digits in the same
  sentence ("회의실 1234번") is dropped; so is an email whose domain you never said.
• Cross-source conflict warnings were inverted — they fired on sources that
  agreed and stayed silent on ones that genuinely disagreed. Conflicts are now
  decided by comparing the actual values (amounts, times, dates — notation-aware,
  so 2pm and 14:00 are the same value) plus a negation check, so "not X, it's
  Y" and purely qualitative disagreements are caught too.

Added

• Muse now tells you why your replies are slow — and it's usually one setting.
  muse doctor measures Ollama's prompt cache directly (it sends the same prefix
  twice and compares Ollama's own timings). Ollama's default splits the cache
  across parallel slots, so the large stable prefix Muse sends every turn — your
  persona, your memory, the tool definitions — never hits the cache, and every
  turn re-pays the whole cost. Measured on a 12B model: an identical prompt costs
  2402ms, 2406ms, 2425ms, 2427ms on the default config; with
  OLLAMA_NUM_PARALLEL=1 it costs 3163ms then 75ms, 69ms, 66ms. Doctor names
  the fix; docs/setup-local-llm.md explains the trade-off.
• Unattended learning is ON by default — but only now that it demonstrably
  works. Every learned strategy still sits on probation until you reinforce it,
  retiring is subtractive-only, the pause switch is honoured, and Muse tells you on
  your own channel when it learned something. Opt out with
  MUSE_SELFLEARN_ENABLED=false.
• pnpm build is 9.3x faster (17.4s → 1.87s): the TypeScript 7 native compiler
  for compilation, and one project-graph build instead of 38 process spawns.
  typescript-eslint and the IDE keep the TypeScript 6 API via Microsoft's official
  compat package, and a new gate keeps the two halves from silently drifting.

---

Early / experimental — macOS only, 0.x pre-release. Local-first: runs on your own Ollama models by default.