By-ear + real-API pass: does the pick actually stop repeating? (spec 03-01 §2.3)

[wine-fall]

What it is

PR #148 shipped two changes aimed at one complaint — the pick task kept landing on the same handful of famous songs (Bon Iver's Holocene was the running example). Half of it is verified at the real seam; half of it could not be run on this machine.

Verified (real brain, real yt-dlp, one MusicProgrammer, no restart): the listener-owned $MURMUR_HOME/music-policy.md steers a live pick. The seeded default policy picked Clairo — Bags; rewriting the file mid-run to "only Cantonese songs released before 2000" made the very next pick 陳慧嫻 — 千千闋歌.

Not verified: the similar_music happy path. No Last.fm key exists on this machine (.env* and the environment both checked), so the only real-endpoint evidence is the failure path: https://ws.audioscrobbler.com/2.0/ accepts our URL shape and rejects a bad key with HTTP 403, surfaced cleanly through LastfmSimilar (src/lastfm.ts:88) — which in the tool becomes {ok: false, error} and costs the pick a turn, not the song. Whether artist.getsimilar / track.getsimilar return usable names, and whether the brain actually widens on them instead of ignoring them, is untested against the live API.

Unblocking it is a free key from https://www.last.fm/api/account/create (no listening account) exported as MURMUR_LASTFM_API_KEY. With no key the tool is absent from the tool list and discovery is exactly its pre-#148 self, so nothing is broken while this sits.

The anti-repeat depth also went 8 → 32 in the same PR. Whether the rotation actually feels wider is a by-ear judgment, and it shares this pass.

Spec

specs/spec03/03-01-brain-harness.md §2.3 (the 2026-08-25 extension) — the two-halves instruction, the key-gated similar_music tool, and acceptance criteria 6 and 7.

Done when

• With MURMUR_LASTFM_API_KEY set, a real run shows similar_music returning real names (the tool result, or the music.* lines in .dev/dev.log).
• At least one pick demonstrably came from a similar_music result rather than a first-guess search query.
• Over a real listening stretch, the rotation feels less repetitive than before — no artist returning every other evening.
• The policy file survives contact with a listener who actually rewrites it: an edit mid-broadcast reaches the next song, and a policy that ignores the built-in advice still produces a track with a spoken announce.

A "no" on any box becomes its own follow-up issue rather than reopening the spec contract.