Should the probe and transfer move onto dl's OpenSSH transport? #424

Question

Should the probe and the transfer move onto dl's own OpenSSH transport too?

This is where the cold path's remaining trip floor gets decided.
#390 put the terminal-borne
dl <spec> -- <cmd> trip on a reusable OpenSSH connection, because dl was
already running OpenSSH there. The probe and the transfer still go through
devpod ssh --command, and they are two of the three cold trips.

#390 measured parity for this in passing, and that measurement is the reason this
ticket exists rather than being assumed either way: same user, same cwd,
--workspace-env values arriving, and 100MB at 4.6s over OpenSSH against 7.4s
through devpod.

Against that sits the objection that has governed this map repeatedly, and it has
to be met rather than noted: this would make two mechanisms for reaching a
container where there is currently one plus one, and this map has consistently
preferred one mechanism over a fast path with a fallback. Weigh in particular:

• What the probe depends on that devpod currently provides. The three-state probe
  (#151, three review rounds)
  reports resolved paths as marked lines and names no state; ProbeResult::parse
  decides on the host. Does moving transport touch that at all, or is it genuinely
  orthogonal?
• The transfer's arch/libc gate, which runs the lent binaries in a staging
  directory and only then moves them into place. It must keep failing closed.
• StderrFilter and interpret, which exist to recover the remote exit
  status from devpod's stderr. What a devpod ssh trip costs per connection, and whether it can be reused #389 measured that remote exit status propagates
  exactly over ssh (3 to 3, 130 to 130), so a direct-ssh path needs neither —
  which means moving these trips would let both be deleted for those paths,
  not merely bypassed. Deleting a mechanism is the strongest argument available
  here under principle 2; check whether it actually holds for these two callers.
• Whether the probe→transfer merge stays refused. It does: Three-state tools probe: only the official claude layout counts as provisioned #151/Collapse the three sequential devpod ssh round trips on the cold path #157 refused it
  on correctness, and nothing here reopens that. Transport is a separate axis
  from trip count.

Blocked on the multiplexed connection landing, since parity has to be judged
against the transport as it will actually be, not as it is today.