Verification: a pull request appears once on the board

The plan in testing-plan.md, executed. Every row is either a result or the reason it does not apply.

Results

#	Type	Result	Evidence
T1	Unit — the join	pass	bun run test -- src/api/model.test.ts — six cases under a pull request is not a work item of its own: the linked, labeled PR renders once; an unclaimed PR keeps its row; a treeless owner's claim is refused for all three treeless loops; a PR with a session record of its own keeps its row and its own tmux target; a self-claim and a mutual two-level claim leave every row standing
T2	Unit — the fold	pass	same file, five cases under what the removed row carried is folded, not dropped: the record and the poll.lastPolledAt age fallback (through sessionTree as well), recent error · PR on the owner's card with the PR's short ref, the needs input chip plus a single Reply entry that replaces the raw awaiting-input row, the human gate not promoted to a chip, and nothing folded for a PR that kept a row of its own
T3	Component / React	pass	bun run test -- src/App.test.tsx — draws a linked pull request once, under its work item and not beside it: #216 is inside Pull requests for loop-lab#214, and no top-level lp-side-row names loop-lab#216
T4	Unit — the service	pass	uv run pytest tests/test_core_attention.py — an armed PR whose only live session is a nested endpoint reports no armed-without-session; a closed nested endpoint still does
T5	Contract (OpenAPI)	n/a	no route, request or response shape touched; docs/api-specs/openapi/ is unchanged, and GET /work-items still serves every portable record (R4.1)
T6	Integration / end-to-end	n/a	no routing, dispatch or registry write path changed; the poller and link_pull_request are untouched. The full Python suite (T10) covers that nothing moved under them
T7	UI / visual	pass	Chromium (Playwright) over the production build in demo mode, 1440×900 @2×, against a fixture that now carries both identities for #216: sidebar-before-duplicate.png — loop-lab#216 as a dead top-level row (dash dot, "2m ago") directly above loop-lab#214, which already nests #216; sidebar-one-row-per-pr.png — the same board with one row for the PR, nested, live, "9m ago"
T8	Accessibility	pass	the nested list is still getByRole("list", { name: "Pull requests for loop-lab#214" }), and the assertion that no top-level row names the PR is read off the rendered links, not the classes alone
T9	Regression	pass	the full vitest suite, including every issue-300 case: labels, treeless loops, the trace tabs, #/item/<ref> and the pre-283 #/sessions/<pr-ref> deep links
T10	Lint / typecheck / build	pass	bun run lint (oxlint --type-aware) clean, tsc --noEmit clean, vite build green; uv run ruff check . clean, ruff format --check clean, uv run pyright the_loop/core/attention.py 0 errors, uv run pytest 2679 passed / 1 skipped; markdownlint clean on every doc touched
T11	Security / abuse case	pass	A1 — the self-claim and mutual-claim cases in T1 prove a record cannot walk the join into removing a row; A3 — T4's closed-endpoint case proves a dead endpoint does not suppress a genuine armed-without-session
T12	Performance	n/a	one extra pass over the sessions array the join already maps, and one nested loop over the list list_attention already walks. No fetch, no hot path

Test counts

• UI: 171 → 183 (+12): six join cases, five fold cases, one component case.
• Service: 2677 → 2679 (+2): the nested-endpoint liveness case and its closed-endpoint counterpart.

No test was weakened, and no existing assertion was edited. Every new test was run against the unfixed code first: the seven UI cases and the liveness case fail there — including issue-300's own "opens the PR's own session from its nested sidebar row", which the fixture's second identity breaks until the join reconciles it.

Notes

• The demo fixture gained the missing half of the real shape: makePullRequestRecord(216), the poll ledger the poller flushes under a labeled PR's own ref. Without it the bundled data could not reproduce the defect, and the screens would keep being evaluated against a board no service produces.
• The T7 screenshots are captured from dist/ built with UI_BASE=/ and served statically, with the demo settings seeded into localStorage before load — the same data the suite asserts against, so the evidence and the assertions describe one board. The "before" shot was taken with the claim check disabled at buildWorkItemViews, nothing else changed.
• uv.lock was left untouched: running uv run rewrites an unrelated typing-extensions marker and the 11.5.0 → 11.6.0 version line that main already carries out of sync. Neither belongs in this diff.