Bug: refreshMessageMetadata decodes a response shape the route never sends — always returns []

[Adron]

Found while implementing #96 (see PR #103), and verified independently against the backend source.

The mismatch

APIClient.refreshMessageMetadata (Services/APIClient.swift:382-391) decodes:

struct Response: Decodable {
    struct Meta: Decodable { let links: [MessageLinkPreview]? }
    let metadata: Meta?          // <- expects a nested "metadata" object
}
return response.metadata?.links ?? []

app/api/messages/[id]/metadata/route.ts returns, at both exit points (:86, :105):

return NextResponse.json(serialize({ links: linkMetadataItems }), …)

Top level links, no metadata wrapper. serialize() normalizes values; it never restructures.
So response.metadata is always nil and the function always returns [] against the live API.

Each entry is also a full LinkMetadataItem — nested metadata object, thumbnail rather than
image, plus platform / fetchStatus / fetchedAt. MessageLinkPreview's flat
{url,title,description,image} shape matches nothing the backend sends; the closest existing type is
CanonicalLinkPreview, which uses imageUrl.

The test is the worrying part

test_refreshMessageMetadata_* passes today. It passes because its fixture was hand-written to the
assumed shape rather than a captured response. So the suite actively asserts the wrong contract.

This is the response-direction twin of the defaultVisibility bug (#46): there, a request key the
route ignored; here, a response shape the route never produces. Both are green-and-wrong.

Build this

• Decode what the route actually returns: top-level links, entries shaped as LinkMetadataItem.
  Prefer reusing LinkMetadataItem over keeping MessageLinkPreview alive — check whether
  MessageLinkPreview has any other consumer first; if not, delete it with its tests.
• Rewrite the test fixture from the route source (or a captured live response), not from the
  Swift type. Add a comment in the test naming the route file it was derived from.
• Confirm the same mistake is not present in the sibling GET on that route (:47-48 returns
  the same { links } shape).

Acceptance criteria

• A fixture matching the real route shape decodes to a non-empty array.
• The old fixture no longer passes.
• feat(messages): apply refreshed link metadata to posted row #103's conversion receives real entries.

Blocks

#103 / #96 are inert until this lands. Does not affect #94, which ignores the response — the POST
still causes the server to fetch and persist metadata.

Files: Services/APIClient.swift:382, Models/Message.swift,
InterlinedListTests/APIClientTests/*.

[Adron]

Implemented in #106 (stack: #106 → #103 → #94). Closing here — further work tracks in the PR.

What the route actually sends, traced to the producer

Read from app/api/messages/[id]/metadata/route.ts, lib/messages/metadata-fetcher.ts,
lib/http/serialize.ts and lib/types/index.ts: { "links": [...] } at the top level from both
exits, entries produced by fetchMultipleLinkMetadata — i.e. exactly LinkMetadataItem
(url, platform, optional nested metadata{thumbnail,title,description,text,type,ogType},
fetchStatus, fetchedAt). Same producer as GET /api/link-metadata, which iOS already decoded
correctly — so the right shape was in the codebase all along, one type away.

Consequences

• feat(messages): apply refreshed link metadata to posted row #103's conversion is deleted, with its 10 tests. The route's entries were already in the feed
  row's shape, so LinkMetadataItem.init(preview:) had nothing to convert. The rest of feat(messages): apply refreshed link metadata to posted row #103 —
  AppDataStore.applyLinkMetadata, Message.linkMetadata becoming var, the ComposeView wiring —
  survives and now receives real entries.
• MessageLinkPreview is deleted. Its only references were the broken decode and feat(messages): apply refreshed link metadata to posted row #103's
  conversion.
• links is now decoded as required, so a future shape change throws instead of silently
  yielding [] again. The publish path's try? still swallows it, so behaviour there is unchanged.
• The old fixture is retained as a test asserting it now throws — the wrong contract is pinned as
  wrong rather than deleted.

Note for whoever merges

#106 undoes part of #103, so the two are worth squashing together at merge rather than landing a
conversion and then removing it in the next commit.

Still blocked for the user-visible outcome

#105 — FeedView renders from a private @State mirror keyed on .count, so the in-place
linkMetadata update still isn't drawn until a full reload. Being fixed separately.