Wave 3 — the words: the comparison in sentences, and the moment the core set is done

[Randolin]

The third wave of docs/feature-plan.md: the copy-heavy work, now that the structure around it is settled.

C2 · The comparison, in sentences

The charts say everything this page knows, and a screen reader can read none of it. The new panel says the same thing in prose: what you both answered the same, where you differ and what each of you actually said, which dealbreakers need a conversation, why fit is two numbers instead of one, and what none of it covers. It sits right after the headline, because it is also what anyone reading under stress reads first.

The phrasing rules are the feature, so they are written at the top of the module and pinned by a spec: describe differences, never grade them; never render a verdict on the relationship; attribute by creature name, because the second person has a name too; always say what was not counted. The spec bans the graded vocabulary outright, and bans minimisers as a pattern (only 40%) rather than as words — "shown only because you both said yes" is precise, and banning it would make the copy worse.

Two category errors turned up by reading the rendered output, not by any assertion:

Identity is not a difference. The first draft led with "On Pronouns, X said they/them and Y said she/her" as a headline gap. Who someone is belongs in their profile and the answer grid, never in a sentence about where two people diverge — for this audience especially. The About section is now excluded from both the agreement and difference sentences.

Care is not a similarity. Excluding identity promoted the give/receive pair into that slot, so the page was about to say "you differ on how care lands best" three lines above "that's a difference in kind, not effort". scores.ts already excludes complement pairs from plain similarity for exactly this reason; the narrative now honours it.

It also stops repeating itself: an item with a dealbreaker sentence no longer gets a plain difference sentence too, which was burying the urgent line under a duller one.

A4 · The core set becomes a moment

coreDone() and a progress ring already existed, so the gap was not tracking — it was that crossing the threshold passed as an inline label quietly changing its wording. It now gets one card, once, that says the part people miss: a finished profile still compares against nothing until someone else's phrase turns up. So it offers the two moves that follow — copy your link, or keep a phrase someone gave you.

It waits for the edit-phrase notice to be cleared first: two loud cards on one screen costs the more important one its attention, and losing the edit phrase is the more important one. Dismissal is per creature and survives a reload, following the notice's own pattern.

B3 · "I've saved it" gets somewhere to go

Half-built already — the acknowledgement checkbox and its gate existed (along with an honest touch: the component tracks whether the clipboard write actually landed, because "I clicked copy" is not evidence the phrase was saved). What it lacked was a destination, because the backup card did not exist when that sentence was written. It does now, so the notice links to it.

That is B3 as this plan worded it ("tick 'I've saved this' next to the backup-card link") rather than the retype challenge listed beside it, which would tax the instant hatch the same plan asks us to protect. The reasoning is recorded in the plan doc rather than left as a silent omission.

Housekeeping

itemLabel moves into core. Five call sites had their own copy in three variants — falling back to the raw id, or to the scale's anchors — and the narrative would have been the sixth.

Verification

Full ladder green: format:check, both typechecks, all four unit suites (23 core files / 4 ui / 6 app / 2 server), build, and e2e. The e2e now also asserts the narrative renders on the demo page, frames the dealbreaker as a conversation, and never frames identity as a difference.

---

Generated by Claude Code