fix(web): conclusion knowledge search with target discovery and correct size pagination

[Poxel2]

Description

Fixes peer-level knowledge search in the web viewer and corrects list pagination across the UI.

Peer knowledge search always returned nothing

The peer page's conclusion search self-scoped the query to observed = <the peer itself>. The conclusions/query endpoint requires both filters.observer and filters.observed, but a peer that holds knowledge about other peers holds nothing about itself — so the self-scoped query was a guaranteed zero rows for exactly the peers where knowledge search matters most.

This PR:

• Adds useConclusionTargetPeers, which walks conclusions/list with an observer_id filter and collects the distinct observed_id values (with row counts) from real stored conclusions. Returns { targets, complete }; complete is false when the walk hit the page cap, so a truncated result is never presented as the full target list.
• Adds a target <select> to the peer page knowledge panel, populated from those discovered targets (masked ids, with counts). A single target from a complete walk is auto-selected; a truncated walk never auto-selects — the user picks explicitly. An explicit user choice always wins over auto-select.
• Removes the implicit self-scope fallback in useQueryPeerConclusions; the query is never fired without a concrete observed target (the server requires both filters).
• Per-peer target state resets when the viewed workspace or peer changes, so a stale target can't silently scope the search wrong.
• Discovery errors are surfaced in the panel and in the picker option state instead of rendering an endless "loading targets…".
• Result rows keep observer→observed, session and timestamp provenance. Identical contents are deduplicated for display only (with a hidden-duplicate count); no data operations.

Pagination used the wrong query parameter

The list endpoints take page and size, but the UI sent page_size everywhere. The server silently ignores page_size and falls back to 50 items per page, so pages beyond the first were never reachable from the UI. All call sites now send size (workspaces, peers, peer sessions, sessions, session messages, conclusions list, dreams, compare views, seed kits, instance discovery).

Testing

• pnpm --filter @openconcho/web test — 122 tests pass, including a new conclusion-targets.test.tsx suite covering: target collection with counts, multi-page walks, complete=false on a cap-truncated walk (40 requests, no walk past the cap), the exact page-count boundary (complete=true when the last walked page is the reported page count), error propagation, the empty-store case, and count/id sort order.
• pnpm --filter @openconcho/web typecheck — clean.
• pnpm --filter @openconcho/web lint — clean (biome, 125 files).
• pnpm --filter @openconcho/web build — succeeds.

Reproduction

1. Open a peer that holds conclusions about other peers and use the knowledge search on the peer page → previously zero results regardless of query; now the observed target is discovered from the peer's actual conclusions and results appear.
2. List any collection with more than 50 items → previously only the first 50 were reachable; now paging works.