fix(openai-history): stop refusals and empty messages from forcing page breaks

[teamchong]

Three independent topics, squashed at request.

Responses planner: responseMessageText gated imageability on an allow-list
of part names (input_text/output_text/text). A refusal part is plain
model-authored prose, but it failed that gate, so the whole message became
non-imageable -- and a non-imageable message is a hard barrier that ends the
collapse run. A single refusal cost a page break. Gate on the payload being a
string instead of on the part's name.

Likewise an empty or whitespace-only message was a barrier despite having
nothing to render; skipping it changes no ordering, so it no longer splits a
run. Items with non-text parts (images, unknown shapes) stay barriers.

Adds barrierTypes to the transform event: the item types that ended a run,
with counts, descending, capped at 8. Diagnostic only -- never affects
routing -- so an incidental page-breaker can be told from an unavoidable one.

Claude profile: BASE_HISTORY was tuned against GPT page geometry, where a page
holds ~660 chars; a Claude page holds ~2750. Re-derived against the same
first-byte latency signal (n=11k) and retuned four settings -- maxImages
32 -> 96, framing full -> compact, factSheetScope per-segment -> combined,
responsesMode pairs -> mixed. 'mixed' matters because 'pairs' only groups
index-contiguous tool rounds and Codex interleaves assistant messages between
them, so every round was landing in its own image. Non-message items remain
hard barriers, so protocol order and open call/output state are unchanged.

restart.sh: replace the duplicated pgrep filter with a list_serving_pids
function, because bash 3.2 (still macOS's /bin/bash) mis-parses a case
pattern's closing paren inside $(...). Also adds --detach, which starts
the proxy under nohup and waits for it to actually bind the port, so
non-interactive callers don't SIGHUP the proxy on exit.

Tests: 5 new barrier-fill cases; 954 pass, tsc clean.