#28433 fix(bench): quote the fake agent's path so fleet agent-mode benches open their chats

The memory bench's agent mode (perf-10x-idle.sh --arg=agents=N, bench_memory.py --fake-agent) has opened no chats on any fleet step. The app splits a harness's args into words the way a shell does (AgentHarnessAddRequest.words). A fleet step's checkout lives under /Library/Application Support/cmux-build-worker/..., so the fake agent's path became two words, the agent exited at once, and the first acp new failed with agent.exited. bench_memory then stops after the launch workspace, and every agents=N sample reads agents=1.

The fix quotes the path with shlex.quote, which words() already understands. The same call in markdown-chat-typing-e2e.py had the same bug and gets the same fix.

Evidence: a fleet step at 6687c32f22bf with agents=5 printed acp new: {"error":{...,"errorKind":"agent.exited","exit":1,...}} and sampled idle-5 agents=1. The same fake agent answers initialize, session/new and a prompt when started with its path as one argument, on Python 3.9 and 3.12.

Performance

Same harness both sides: scripts/measure/perf-10x-idle.sh --arg=counts=1 --arg=agents=N --arg=measure=… on one fleet Mac, both runs on the same worker (fresh-home tagged build). This PR changes no app code; the rows show the harness now measures what it is asked to.

metric	before	after	delta	harness or run link
workspaces in the idle-20 sample (agents=20)	1	20	+19	before: cmux-ci step 7aa4a5e6f2cb1f3b6413623b (9f45a27f); after: step 5d283c1b76d4e8112ee62c50 (9d2e8cc8)
agent chats opened by --fake-agent	0 (acp new: agent.exited)	19	+19	same steps; before also step 81f44712f605894f573b79a1 (6687c32f, agents=5)
cef-renderer processes at idle-20	5	24	+19	same steps
cef-renderer footprint at idle-20	290.7 MB	1840.0 MB	+1549.3 MB	same steps

No capture: harness scripts only, nothing visible.

Agent session: agent session

🤖 Generated with Claude Code

Agent session: agent session

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

---

Summary by cubic

Fixes agent-mode memory benches so fleet steps actually open chats instead of sampling agents=1.

• The app splits a harness's args into shell-style words, and fleet checkouts live under a path with a space, so the unquoted fake-agent path became two words and the agent exited at once.
• Quotes the path with shlex.quote (which the word splitting already understands) in bench_memory.py and applies the same fix in markdown-chat-typing-e2e.py.
• Harness-only change; no app code or perf numbers are affected.