ci: show why the refresh-cost step failed

[arzafran]

Refs #183.

What happened

#223 added a dispatch-only step to measure the text-input refresh cost in-process. I dispatched four runs (two on main, two on the fix branch) and all four failed with a single line:

FAIL: Socket closed while waiting for response

That tells you the app died or refused the connection. It does not tell you why. The app's own log is written to /tmp/programa-ci-refresh-cost.log on the runner and never surfaced, so four runs produced no diagnosis.

I guessed twice at the cause — a DispatchQueue.main.sync deadlock, then an assert(Thread.isMainThread) trip from an off-main sendEvent — and both were wrong on inspection (main.sync is the established pattern at eight other sites in that file, and simulateShortcut does hop to main). Rather than guess a third time, this makes the next run answer the question.

What this adds

On failure the step now reports, before exiting non-zero:

• whether a Programa DEV process is still alive
• whether the socket exists at the expected path
• the last 80 lines of the app log

Test plan

• YAML parses; step body passes bash -n
• CI green (the step is dispatch-only, so it does not run on this PR)
• After merge: re-dispatch and read the actual cause