feat!: try a key's badges in order and stop at the first match

[GraemeF]

A badge's match is how a config says which values it wants. Until now a pattern that failed to read the value was treated as a badge falling short, and the row said so. That fires on every value a shared list was never meant to read. The live delivery_pr badge reads owner/repo#N, and bdi-pj92 records four closed beads carrying a URL or a bare number, so the bead window reported a reference lost where nothing had been promised about the shape.

Badges on one key are now a chain. Each is tried in config order, the first to read the value draws, and the ones below it never run. A value no entry for the key reads draws nothing and reports nothing, whether or not those entries carry a link. Graeme chose first-match over every-match because ordering then carries meaning. A reader who does want every value of a key writes a last entry with a permissive pattern, which the schema already allowed, so nothing here adds a key or a field.

Breaking. Undrawn::Badge and the row no badge for <key>: no pattern reads this value are gone. A config with two entries on one key whose patterns both read a value drew both and now draws the first. Undrawn::Link and Undrawn::Short are untouched: those are a badge that read the value and then could not keep a promise its config made, which is the case a reader cannot see for themselves.

Shadowing goes with it. Config::badges_for_project used to replace every shared entry for a key a project named. It did that to stop a project's entry and a shared entry both drawing for one value, and first-match now does that on its own. All the replacing still achieved was losing the shared shapes: a project naming delivery_pr to add a bare number stopped reading the qualified form. So a project's entries for a key now stand where the first shared entry stood, and the shared entries follow them. The project wins a value by being tried first, and a value it does not read falls through.

That gives up suppression. A project can no longer silence a shared badge on a key it names. No config wanted to, but one test asserted it, and its doc comment argued for it: shadowing half a key, it said, would leave a project drawing the shared wording for values it did not name. Ordering is now how a reader asks for exactly that, so the worry is answered rather than dropped.

Two tests went with the retired variant rather than being rewritten around it. The model test asserting a report for an unreadable value is the retired behaviour itself. The row test asserting an anomaly's key reaches the row's notes has a sibling asserting exactly that for Undrawn::Link, which carries the property alone now.

Every behaviour claim here was checked against a deliberately weakened tree rather than left to read as passing. Disabling the fall-through guard makes the new pty test fail on the raw value. Making a key's badge read its value and then fail its link puts that key's name back on the screen, which is what that test's silence assertion looks for the absence of. Restoring shadowing makes the new snapshot test fail on the bead the shared entry reads.