M5: Tauri app alpha — Shelf, Story, Compare, watcher

[sepehrsadri]

Tracking issue for M5 (docs/ROADMAP.md "Now: the 0.0 pilot"). Needs M3 (+M4 for
waveforms). Gated on M2.5 (#15) — if that measurement comes back with >50% of real
Logic saves showing no visible structural change, this milestone waits in favor of the
M2 volume-fader stretch goal.

Scope

Menu-bar-first (ActivationPolicy::Accessory, tauri-plugin-positioner), app/src-tauri
as its own Cargo workspace with path-deps on crates/* (root workspace stays
crates/*-only forever — see the M0 guardrail).

Ships:

• Shelf — one card per discovered project, artwork from Logic's own
  WindowImage.jpg.
• Story — reverse-chron timeline, session grouping (>45 min gap), provenance badges,
  zero-change saves auto-folded.
• Compare — grouped plain sentences, deterministic order, units translated, IDs never
  shown.
• Bounce drop-zone (M4's null-diff, works for any DAW).
• "Copy as text" — promoted into M5, not deferred with the M6 share HTML. This is 0.0's
  entire sharing story.
• Trust panel (watched folders, "nothing leaves this Mac", opt-in research report).
• FSEvents watcher — always ingests byte-different saves (see M3's retention rule);
  semantic verdict only labels.

Non-negotiable copy rules

The capability caveat rides on every aggregated surface, not just the Compare footer —
the menu-bar quiet state reads "no changes Wit can see this week", never "nothing new."
Logic's empty verdict is grey and scoped ("Wit can't yet see knob and fader moves in
Logic"), never a green check and never "no musical change" (that string is Ableton
Semantic-tier only, and Ableton isn't in 0.0).

First-run defaults to oldest-vs-newest aggregate, never an adjacent pair, for the
first compare. Duplicate-audio hashing runs later as a background job, never inside
onboarding.

just demo-library builds a synthetic ~/Music tree so first-run, timeline, and watcher
are demoable on any machine, not just the machine with real Logic projects on it.

Exit criteria

Installs on a second Mac, points at the demo library and a copy of a real Logic
folder, reads a diff, clicks Reveal.

[sepehrsadri]

The M2.5 gate is cleared — this milestone is unblocked.

This issue was gated on #15: "if that measurement comes back with >50% of real Logic saves showing no visible structural change, this milestone waits in favor of the M2 volume-fader stretch goal."

Measured at the scale #15 specified (32 projects, 28 GB, 100 consecutive save pairs, 2026-08-16): 33.0% show no visible structural change, well under the 50% threshold. Full finding in PR #37 / docs/EXPERIMENTS.md §11.

Two things worth carrying into the M5 build from that data:

1. A third of saves will still read as empty. The scoped grey copy in this issue's "non-negotiable copy rules" is not a nicety — it is what a third of the timeline will actually say. Auto-folding zero-change saves in Story matters proportionally more than it looked.
2. The common non-empty case is small. The distribution is long-tailed, not bimodal: the mode above zero is a single change, and 26 of 67 changed pairs sit in the 1–7 band. Compare should be designed for "a handful of sentences", with the 235-change outlier as the exception it is — not the reverse.

The volume-fader stretch goal is not dropped, just no longer sequenced ahead of this: a save that moved only a fader is inside that 33%.