test(e2e): enrich LR-DIAG dump with first/last 5 reqs + roles/path/system markers

After last CI run we saw 1770 mock requests in turn 4 hang, all showing
the SAME last assistant message in the dump. That means our prior dump
(messages.at(-1) only) couldn't distinguish:
  - Whether those were historian subagent requests vs main-agent requests
  - Whether request path varies (e.g. /messages vs /count_tokens)
  - The actual roles sequence in messages array (could be 1 user + many
    assistant turns, vs 200 alternating)
  - Whether opencode was retrying the same identical request

Enriched dump now shows for first 5 + last 5 of the dropped requests:
  - path + msgs.length + model + histReq/mcReq classification
  - full roles list (e.g. 'user,assistant,user,assistant,...')
  - lastRole + lastContent preview

This should pinpoint whether the 1770-request loop is:
  (a) Plugin spawning thousands of historian/dreamer subagent calls
  (b) OpenCode internally retrying after each failure
  (c) Some other request type (count_tokens, token estimation)
  (d) Genuine assistant→user→assistant loop without termination