CL-6436: fix stale mention-fan-out chat e2e (walking-skeleton red)

[TheGreatAxios]

Diagnosis: stale test, not an endpoint bug

chat e2e > mention fan-out drives the mentioned run 404s on every recent
main run:

list run events for run_...: expected HTTP 200, got 404 "Run not found"

Traced to the endpoint's actual, correct behavior:

• d5c451d3 "Chat: delete the workbench anchor machinery (CL-6330)"
  (same day as CL-6451, landed first) states plainly: "A workbench is
  data. Creating one mints a child tenant and writes settings rows — no
  host workflow, no asset, no run, no deploy."
• The old test (from ad1eb23e, the day before) mints a second bare
  kind: "workbench" room, adds its raw address as a mentionable
  participant of the first room, then polls
  GET /workflows/runs/:id/events using the second workbench's own id
  as the run id.
• Since CL-6330, a bare workbench has no workflow_run row at all — its
  id is shaped like a run id (generateId("workflowRun")) but is only
  ever used as a tenant id. findRoutableById (vendored
  hub-sessions/src/hub-session-lookups.ts) correctly 404s: there is
  nothing to find.
• Confirmed this via git log -p — the CL-6451 PR's own later test
  ("inviting the echo agent...") comments that the earlier "mention
  fan-out" test proves "the same fan-out pattern... for a
  workbench-to-workbench mention," a claim that stopped being true the
  moment CL-6330 landed later that same day. Nobody re-validated the
  earlier test against the new no-anchor model.

Only an agent participant (invited via definitionId, through
kind: "chat" or POST .../invite) is backed by a real workflow_run
(mintFoldedRun). Mentioning a bare workbench was never going to have a
run to query.

Fix

Rewrote the test to mention a real resident agent instead of a bare
workbench, and strengthened it into an actual CL-6451 assertion (one room
participant, one live run — never a sibling):

• Creates a kind: "chat" workbench with the echo definition (a
  genuine invited agent, so a real workflow_run exists).
• Sends two separate @echo mentions and asserts the same run's
  event log progresses past its own watermark both times.
• Asserts the room still lists exactly one participant (the same
  address) after the second mention — proving no sibling run was
  minted, which is the actual CL-6451 contract this test was meant to
  guard.

Also fixed the file's now-stale header comment (still described the
pre-CL-6330 "anchor instance" model) and dropped the now-unused
formatRunAddress import.

Verified

• bunx eslint scripts/e2e/chat.test.ts — clean
• bunx prettier --check scripts/e2e/chat.test.ts — clean
• bunx tsc --noEmit — no errors in this file
• Git-archaeology of d5c451d3 (CL-6330), ad1eb23e, and PR CL-6451: one room participant, one live run — command paths reuse resident agents; stale turns fail #181
  (CL-6451/be0c9544) to establish the ordering and confirm the anchor
  model was genuinely deleted, not just refactored

NOT verified — needs a reviewer with a live stack

• Did not boot the walking-skeleton stack myself (machine is
  memory-constrained and another lane holds a live stack for Ollama
  testing right now) — so this fix is unconfirmed against a real hub +
  sidecar + scratch DB run. A reviewer should run this file's chat e2e
  suite (or let CI's walking-skeleton job do it) before merging.
• Did not audit every other e2e file for the same stale-anchor-model
  assumption; this PR only fixes the one test named in CL-6436.

Do not merge — flagging for peer review per the task's own
instruction.

Fixes CL-6436