Explain v1: a model-free Compare view for a Similar card #236

Part of #235.

What

An Explain action on every Similar & unlinked card (beside Why?, and in the card menu for the keyboard). It opens a Compare mode in the center pane, the way Graph mode does: the open note (the anchor) against the candidate, built only from stored data.

┌ Volcano notes  ↔  Weekly journal 2024-W12 ─────────────── ●●● 2.8σ · #2 of 10 ┐
│ One section matches strongly; the rest of the note is about other things.      │
│ (Whole-note rank #87 → best-passage rank #2)                                   │
│                                                                                │
│ Where it sits among the 200 unlinked notes compared                            │
│   · ·· ····:::::||||||||:::···· ·    ·   │1.96 ●●○   │2.52 ●●●   ▲ this card   │
│                                                                                │
│ Best pairs                                                                     │
│ 1 ●●● Lava fields ↔ Thursday     2 ●○○ Eruptions ↔ Thursday    3 ...           │
│ ┌ Your note › Lava fields ───────┐  ┌ Journal › Thursday ────────────────────┐ │
│ │ ...basalt flows cool into ropy │  │ ...walked the pahoehoe crust at dawn,  │ │
│ └────────────────────────────────┘  └────────────────────────────────────────┘ │
│ Links: not linked. Shared neighbors: Hawaii trip                               │
│ [Open]  [Ask chat why ▸]                                                       │
└────────────────────────────────────────────────────────────────────────────────┘

v1 scope

1. Facts line: rank in the served list, strength band, z; the candidate's whole-note (centroid) rank beside its best-passage rank.
2. Summary label from fixed, documented rules (no model):
   • Identical text: the winning pair shares a text_hash (the same passage is in both notes, e.g. a template).
   • One section matches: best-passage rank far better than whole-note rank.
   • Broadly similar: several candidate passages pair at or above the ●●○ landmark.
   • otherwise Nearest in meaning.
     The labels grade nothing and gate nothing; their rules are written down in the view type's docs.
3. Where-it-sits strip: every scored candidate's z as a dot, the two band landmarks, this card marked. Shows "strong compared with what", and the Discovery floor: the z rule assumes an unrelated tail that a single-domain vault does not have — 16 of 17 notes serve nothing #196 compression at a glance.
4. Best pairs: the top passage pairs (discover::passage_pairs), each with both passages' heading path and text, each graded in the dot bands.
5. Link context: directly linked (then it would not be listed), shared neighbors.
6. Hand-off: Ask chat why runs today's Why? on the same pair.

Engine

• discover::candidates also reports the population behind the dots (size, mean, sd of best-passage squared distances) and each row's stage-1 (centroid) rank. Row order and z are unchanged, and a test pins that.
• New façade op Vault::explain_similar(anchor, candidate, limit) -> SimilarExplainView. Works for any pair: a candidate outside the served list still gets its facts (rank past the list, or excluded as linked).
• CLI: b2 similar NOTE --explain OTHER [--json]. b2 similar's own --json is unchanged.
• Desktop: one explain_similar command reusing the view type.

Out of scope for v1 (see #235 children)

Match map heatmap, shared words, the "why isn't X here?" picker, jumping to the passage inside the note (#232).

As built (PR #242)

Two decisions changed during the build. The spec above is kept as written; this is what shipped and why.

1. discover::candidates is unchanged. The population statistics and the stage-1 (whole-note) rank are reported by a new discover::explain instead. Both read one shared field computation, so the card and its explanation still can't disagree, and b2 similar --json and the Similar list keep their exact shape and order.
2. The summary labels count passages at the ●●○ mark instead of using a rank gap. A rank-gap rule ("whole-note rank far worse than best-passage rank") needed a new, uncalibrated threshold. Counting candidate passages whose pair reaches the existing 1.96σ mark reuses a measured number. The rules, in order: Identical text; no label when ungraded; then No passage stands out (none reach the mark), A clear match (its only passage does), One section matches (one of several), Broadly similar (at least half), A few sections match (otherwise). The whole-note vs best-passage ranks are still shown as a fact line whenever they differ. Documented in ui/src/explain.ts.

Added during review: a σ axis with a key and a "?" account on the strip; same-titled notes named by path; the card count capped at what the list shows; the CLI says when pairs are ungraded; docs say explanations agree with the list at the same --limit.