feat: three layers, hover highlighting, topic filters, and how to distil a persona

[khaneight]

All four things asked for.

The graph is three layers, and the middle one is the persona

ring
Source material	documents the author wrote
Persona	cited traits distilled from that writing
The clone's work	articles the archive produced

LAYERS is published in the bundle with names and descriptions, so the page
cannot invent its own account of what the archive is made of.

An affirmed trait is now a node, edged down to every document it was read
out of and up to every article written from it. That chain is the claim the
whole picture makes: a sentence in a generated essay can be followed back
through the trait that produced it to the passage in the corpus that supports
it. proposed traits are absent — drawing an unconfirmed reading of somebody in
the middle of the picture asserts it to every reader before they have seen it.

Every article sits in the work layer whatever its origin. Depth answers
what kind of thing is this; origin and extrapolated answer how was it
written. Splitting rings by origin would put a compiled and a machine-written
article at different depths while the legend claims depth means layer.

Hover, and filters

Hovering lights a node's edges in its own colour, holds its neighbours bright
and dims the rest. Filters are built from the archive's own vocabulary — domains
and tags, with counts, any number selectable. Traits have no topic of their own,
so they stay while anything they touch does.

Trait labels on the graph are ids rather than claims, since a claim is a whole
sentence; the claim is still the title in the reader, where there is room.

value joins the trait kinds

Ordered surface inwards: style, pattern, principle, value, belief.

A principle says do this; a value says this matters, and the same value
produces different rules in different settings. Keeping them apart is what lets
the clone extend a view rather than only repeat its conclusions.

/sentinel-clone gains a method

One pass per kind, with the question that finds each and the test for whether
you actually have one:

kind	the question	the test
style	how does this read?	could a reader tell their paragraph from someone else's?
pattern	what move do they keep making?	can you point at it twice?
principle	what rule are they applying?	can you state it as do this?
value	what do they treat as worth having?	does it survive when the rule it produced would not?
belief	what do they hold to be the case?	could it be false?

Plus calibration for confidence, and an explicit instruction not to fill the
table — a trait invented to complete the set is one the user then spends time
rejecting.

539 tests (+3), clippy and fmt clean. CLAUDE.md compressed back under budget
rather than raising it again, as its own constant instructs.

🤖 Generated with Claude Code