cap feedback: read community feedback on a dataset's labels

[dosumis]

CAP dataset pages carry a /feedback tab where readers score a label and can attach a typed explanation -- a comment, a refine naming a replacement ontology term, or a split/merge proposal with markers. None of it is reachable programmatically today, and it matters: of the first 63 findings in one HCA annotation review, six had already been filed there by other people, and none had reached the labels or CAP's exported ontology report.

There is no API. The GraphQL endpoint safelists query bodies and the page's DatasetFeedbacks operation runs server-side, so nothing leaves the browser to replay. The result is embedded in the page's Next.js payload, which is the same route labelsets already takes -- so feedback.py mirrors labelsets.py: a pure parse_feedback(html), and a fetch_feedback(..., fetch_page=None) that is injectable and therefore testable without the network.

Two deliberate behaviours:

• fetch_feedback RAISES when a page yields no labels, after retrying. CAP intermittently serves the page shell without its payload, and an empty parse is indistinguishable from "no feedback" unless you refuse to guess. Reporting zero feedback for a dataset that has some is the failure that costs you; an earlier scan silently did exactly that to a dataset with 14 feedback labels.
• Attribution is passed through -- displayName, orcidId, institution -- because feedback is signed work. Email is never emitted, even though the payload carries it. A test asserts that.

Also moves the page fetcher from labelsets into transport as fetch_page(), rather than duplicating it: two structures now need it, so it is a transport concern.

The new SKILL.md pins the branch, not v0.3.0, because cap feedback does not exist in that tag; change to @v0.4.0 when the release is cut.