fix(board): spend the search limit per relation, batch the checks

[tensorcopy]

Why. The search limit was one shared budget across every relation, so a full review queue — whose items are the most recently updated — took the whole list and pushed the viewer's own work off a board that exists to show it. Checks assumed the merged list always fit a single nodes(ids:) request.

What.

• The limit is now spent per relation: each relation keeps its own newest limit items, and the union keeps whatever any relation kept.
• attachChecks asks in batches of 100 ids, so a longer board keeps its checks; one failed batch costs its own pills, not every other batch's.
• The lockfile's version fields are synced with package.json (they still said 0.5.0), as a chore commit of their own.

How. mergeBucketResults in server/board/buckets.ts collects the kept ids per relation after the merge, so an item any relation kept survives; server/board/checks.ts walks the id list in CHECKS_BATCH slices.

Test. server/board/buckets.test.ts gains the per-relation budget cases. Typecheck and the suite pass at this tip: 56 tests. Every later commit in this stack also builds and tests on its own, checked commit by commit.

[tensorcopy]

Hi @alysnnix! 👋

I've been running a fork of this plugin for a few weeks and accumulated some changes I'd like to give back. They're cut into a small series so each piece is reviewable on its own:

• fix(board): spend the search limit per relation, batch the checks #6 — two board fixes: the search limit is spent per relation instead of one shared budget, and the checks query is batched. Plus a one-line lockfile version sync.
• feat: open the original chat from a pull request #7 — Open original chat: a pull request the board shows is linked to the first agent created in the workspace tracking it, with an action on the row and in the panel. Independent of fix(board): spend the search limit per relation, batch the checks #6 — either order works.
• feat: aggregate all authenticated GitHub hosts #8 (draft) — every authenticated GitHub host swept together, each read and action routed to the host that owns the item. The draft flag is only to keep the merge order; the code is final, and the note at the top of the PR says which commits to read.
• feat: group the board by host and paint it at once #9 (draft) — one section per host, the board painted from the daemon cache and refreshed behind it, and attachments proxied on authenticated hosts with each host's own token.

Everything passes typecheck, lint and the test suite on every commit, and each PR body follows the repo's conventions as I read them. No rush at all — and happy to rework, re-cut, or drop anything that doesn't fit where the project is going.

Thanks for the plugin, Alysson — it's been a great base to build on. 🙂