docs(roadmap): the login the catalogue asks for, and an order to read the lines by

[wine-fall]

What this is

ROADMAP.md gains a sixth line and a priority column, and its
Tracked-as cells catch up with what has landed since it was written.

The new line 5 — an opt-in login for an auth-gated catalogue (P1)

Not a new want: specs/DESIGN.md §5 already excluded both sources by name and
said why — NetEase Cloud Music has the best Chinese catalogue but only
unofficial APIs, needs a login cookie, and gates its good tracks behind VIP;
Spotify has no clean no-app-no-membership path, and headless librespot needs
Premium. specs/spec03/03-01-brain-harness.md deferred it as a single work
item and named cliamp as the credential
reference — the auth flow only, never its user-picks interaction model.

What is new is the lazy road in: yt-dlp — already murmur's only provider —
ships netease:song / playlist / singer / djradio extractors and takes
--cookies FILE / --cookies-from-browser (checked against yt-dlp
2026.08.19). So the NetEase half is not a second provider; it is a credential
reaching the provider murmur already runs, plus URL-shaped candidate sources
beside the open-ended search. MusicProvider (src/contracts.ts:175) is
search + resolve with no notion of a source that can fail on auth, and
src/app.ts:248 constructs YtDlpMusicProvider directly — so an expired cookie
would surface today as one more dead stream probe, which is the exact confusion
line 3 exists to remove. Spotify gets the seam and an honest refusal.

Opt-in, never a shipped default (DESIGN §3.7's personal-experiment tier): the
default install stays login-free yt-dlp.

The P column

The row numbers were the order, and the order has moved. P0 = line 0, because
every by-ear judgement above it is worthless while the listener's first typed
line disappears. P1 = lines 1, 2 and 5 — the ones that change what murmur is.
P2 = line 3 (reliability). P3 = line 4 (distribution and not rotting), with a
note that its eval half (#98) becomes P2 the moment lines 1/2 land, since both
are entirely stochastic behaviour.

Line 0 gets a status check

Recorded rather than left for the next session to re-derive: no fix has
landed for the dropped first line — the settled guard in lineReader
predates the 2026-08-25 repro (it came in on 08-18/08-19) and LineQueue has
not changed since. A stub plain-mode run with a pre-seeded persona echoed and
acted on the first typed line, but that path runs with no harness, so neither
the crash-report offer nor the setup conversation opens a pre-broadcast reader
— it therefore neither reproduces the bug nor clears it. The suspected seam
also reads clean today (every read() is awaited; settled is set in the
race's own finally), so the suspected cause is marked unconfirmed and the
next attempt should start from the repro conditions, not from that seam.

One test fix rides along

CI came back red on a docs-only branch: test/memory-fold.test.ts:312 reopened
PersistentMemoryStore with the real clock while the rows under it were
written on a frozen 2026-09-01, so the 48h recent window in load() dropped
them for age. It passed the day it landed (#196) and went red on 2026-09-04
with nothing changed — a calendar failure, not a regression, and it blocks
every PR until it is fixed. The reopen now reads the same clock as the writes,
which is what that assertion was ever about: the row survived forget, not the
window.

Verification

Docs, plus that one test line. pre-commit run --files ROADMAP.md → source-language gate passed;
tsc / oxlint / path-governance had no files to check. Every factual claim in
the new text was checked against the tree: yt-dlp --list-extractors and
yt-dlp --help on 2026.08.19 for the extractors and cookie flags, git blame

• git log for the settled guard's date and LineQueue's history, and the
  two code references (src/contracts.ts:175, src/app.ts:248) read directly.

Full local gate after the test fix: npx vitest run --dir test 1366 passed / 0
failed, tsc --noEmit clean, oxlint clean.

Peer review (codex gpt-5.6-sol): 6 findings, 6 applied, 0 dismissed —
verdict was request changes, all six on claims the document overstated, and
all six now corrected in a follow-up commit:

• speaking over a ducked track is not what is missing on line 1 — the
  lead-in (feat(music): the way into a track, not a label on it [spec 03-02] #199) and the coda (feat(director): a coda — the beat that knows the song happened [spec 04] #200) both do it today; Interject a talk beat over a ducked song: the engine ducks, the director never asks [spec 03-02 §3.5] #163 is the autonomous
  mid-track beat, and the line and its table cell now say that;
• a yt-dlp auth failure comes back through provider.resolve and is caught by
  submit_pick as "pick another" — it never reaches the stream probe, so an
  expired cookie reads as a silent run of rejected candidates, not "one more
  dead probe";
• a free Spotify account does have a stream (the desktop app, ads and
  on-demand limits, per DESIGN §5) — the refusal is that murmur cannot conduct
  it, the external-player duck path being out of scope in spec 03-02;
• line 5 now says outright that it does not reopen DESIGN §8's v1 exclusion;
• the crash-report offer landed 08-31, after the 08-25 repro, so it cannot
  have eaten that first line — the repro conditions are the August boot state,
  and the offer is a separate new reader worth its own check;
• feat(tui): the guide gets a real composer, and your line is in the log at once [spec 10] #187 did touch LineQueue (hasReader() + the IpcHost echo flow), so the
  "unchanged" claim narrows to the take/peek semantics the plain path uses.

It also confirmed independently that yt-dlp 2026.08.19 lists the cited NetEase
extractors and accepts both cookie options, and that the two code references
match.

AI coding brief

Original request. The user asked what is on the roadmap and what is done —
which surfaced that ROADMAP.md had gone stale (line 2 built and unmerged,
line 1 half-landed). They then asked to add the NetEase/Spotify login they
remembered from the original specs, mark it P1, and assign priorities to the
existing lines. The why: the roadmap is the document the README points at, and
an order that only lives in a conversation is not an order.

Manual interventions. The user set line 5's priority (P1) themselves and
left the rest of the priorities to analysis. They also pushed back on line 0 —
"I thought that was fixed" — which turned into the status-check paragraph after
git blame showed the guard predates the repro and a stub run turned out to
prove nothing either way.

Retro. The roadmap and the issue tracker had already drifted apart in four
days; reading gh issue list + gh pr list before answering "what is done"
was what caught it, and doing that first would have made the first answer
correct rather than the second. Verifying the yt-dlp extractor claim before
writing it into the doc (one --list-extractors call) is what turned line 5
from "build an auth provider" into "pass a cookie to the provider we run" —
worth the 20 seconds, since it changes the size of the work.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Sx3J45TyoScFDneRF7qX14