df publish: key ui/<sha>/ artifact dir by routeId/kind to prevent cross-attribution across multiple artifact-bearing routes

[alien8d]

Problem

collectPublishArtifacts (packages/cli/src/evidence/publish.ts) reads route
file artifacts from agent-reviews/quality-gates/ui/<sha>/, a path keyed by
<sha> only — not by routeId or evidenceKind. With a single
artifact-bearing route (the reality today) this is correct. With more than one
artifact-bearing route armed for the same SHA, every such route reads the same
directory, so the same files are cross-attributed to each route in the
PublishedEvidence manifest.

This was latent while the upload gate was hard-coded to evidenceKind === "playwright" (effectively one route). #258 broadened the gate to a membership
test over all non-none artifact-bearing kinds (playwright, migration, terraform, test, docker), which widens the surface where a multi-route change
could cross-attribute.

Impact

A multi-route verification run can publish screenshots / test-results / docker /
plan artifacts under the wrong route, making the manifest misstate which proof
supports which objective — affecting downstream worker joins and dashboard proof
review. (Flagged [high]/contracts by the local codex chief-engineer critic on the
#258 PR.)

Required fix

Make artifact discovery route-scoped before broadening further:

• Partition producer output by routeId (and/or evidenceKind), e.g. write to
  agent-reviews/quality-gates/ui/<sha>/<routeId>/..., and have
  readEvidenceArtifacts read the per-route subtree (with a back-compat fallback
  to the flat ui/<sha>/ path for the existing playwright producer).
• This touches the producer contract (skills/verify/producer/*), not just
  the publish reader — which is why it was deliberately scoped out of Generalize evidenceKind beyond 'playwright' for artifact upload (publish.js:120) #258.
• Add a regression test with two artifact-bearing routes in one gate JSON
  proving their artifacts do not cross-attach.

Context

Deferred from #258 (the brief scoped this caveat as a known follow-up). The #258
change documents the limitation inline in readEvidenceArtifacts' doc comment.

[alien8d]

This now subsumes #258. The evidenceKind generalization (#258) and this artifact-dir keying must ship together: generalizing the artifact read to all kinds without keying the dir cross-attributes artifacts across routes (codex flagged it [high] on the #258 prototype).

Scope of the combined change:

• Producer side: the route scripts that write agent-reviews/quality-gates/ui/<sha>/… need a routeId (or kind) segment in the path.
• Consumer side: packages/cli/src/evidence/publish.ts reads that dir — generalize the gate from === "playwright" to a membership test over artifact-bearing EVIDENCE_KINDS, keyed by the same segment.
• A read-side membership-test prototype already exists: Generalize evidenceKind beyond 'playwright' for artifact upload (publish.js:120) #258 commit 18bfb04 (df/258-evidencekind-generalize).

Consumer-contract surface → version bump + docs/CONSUMER-ADOPTION.md update.