test: yjs suites and bench scenarios fit raak's attachment policy, placement and task edits

[flipvh]

After the 0.18.0 sync, pnpm test had two failing yjs integration tests and pnpm bench --all failed four of its eight scenarios. None of it was a product regression: the tests and the bench payloads did not fit raak's policy and API. This PR makes them fit. No product code changes.

Change

• Yjs integration tests. permissions.test.ts and revocation.test.ts expect that an organization member may update their own attachment; raak's policy gives a member update: 0. The backend suites already declare that assumption with assumeMemberAttachmentPolicy. The helper moves to shared/src/testing/member-policy.ts, backend/tests/security/helpers.ts re-exports it so the seven backend suites that use it are untouched, and both yjs suites call it with { read: 1, update: 'own', delete: 'own' }.
• attachment-churn (new in 0.18.0, never passed here). Its create payload carried no projectId, which raak requires, so every create answered 400 and every delete 403. A batch now lands in a seeded bench project, with a fork marker as in attachment.bench.ts.
• task-edit and cdc-throughput (raak-owned, both use processors/task-edits.ts). The assign edit sent a field timestamp for assignedTo, a set delta, which the API refuses: 25% of the requests answered 400. It could also name the same user twice in one delta, which the API refuses as well: rare, but enough to fail a --short run. It now sends no timestamp and distinct users.

Verification

• pnpm check and pnpm unused pass.
• pnpm test: 4909 passed, 0 failed (before: 2 failed). A dev stack was up, so the bench smoke test ran all eight scenarios short.
• pnpm bench --all, then the two scenarios a dev-stack restart interrupted, each on its own:

Scenario	Before	After	Req/s	p95 (ms)
attachment-churn	2500 of 2500 rejected	passed	107	32.8
attachment-edit	failed once on old counters, see below	passed	789	1064.4
cdc-throughput	25% rejected	passed	631	77.5
get-me	passed	passed	1330	54.1
page-load	passed	passed	876	18
sse-fanout	passed	passed	16	24.8
task-edit	25% rejected	passed	469	26.8
yjs-typing	passed	passed

The latencies come from a laptop that also ran the stack and other work: read them as rough.

Notes for the reviewer

• Not fixed here, nothing to fix: attachment-edit failed its first run on the CDC worker's books check. The first verify after 0.18.0 found 202 counters the old worker had left (200 are the new e:c:h:workspace and e:c:h:project keys, 2 are label usage counts from a soft-deleted task) and rebuilt them; every later verify found 0 differences. Production will see the same once, at the first nightly verify (03:00 UTC) after the 0.18.0 deploy: one wrong_books incident, and connected clients refetch.
• cdc-throughput still does not check the worker's record. Its clean run had 25,000 accepted edits and 22,935 recorded activities: an edit that changes nothing writes no row, so counting accepted responses would fail that check.
• Both the helper's place and a home for bench attachments belong upstream; they are noted as cella feedback.

🤖 Generated with Claude Code