test(ssh): state mux explicitly so the ControlMaster tests are platform-stable

[dizhaky]

Correcting myself first

Across #174–#178 I recorded that "Windows' no-mux SSH path has no test coverage at all." That was false.

ssh-connection.test.ts has seven dedicated no-mux: tests, covering every one of the five mux-conditional branches in ssh-connection.ts:

• argument construction carries no ControlMaster/ControlPath/ControlPersist
• open() verifies with a one-shot exec, no -M
• open() classifies auth failure
• forward() spawns the persistent -N -L child; cancelForward + close kill it
• fail-fast when the tunnel child dies
• an unrelated listener can't mask a delayed bind failure
• tunnel death after readiness makes the connection unhealthy

I inferred the gap from watching the suite fail on Windows and never opened the file to check. It was cheap to check.

What's actually wrong is close to the inverse

All eight Windows failures were in mux tests that never set mux, so they inherited:

this._mux = opts.mux ?? process.platform !== 'win32'

On Windows that resolves false — the tests silently became no-mux connections, and their ControlMaster assertions couldn't hold. The suite wasn't missing no-mux coverage; its mux coverage was quietly evaporating on the one platform it was never run on.

The change

mux: true stated on all sixteen ControlMaster constructions. controlDir is the exact marker — it's meaningless without a master — so the edit is mechanical and complete rather than judged case-by-case.

Provably a no-op on Linux. The only value changed is opts.mux, and both undefined ?? true and true ?? true resolve to true there. A regression on the platform CI runs today isn't possible from this change. Simulated across all three platforms:

  linux   mux tests: before=true  after=true    no-mux tests: false
  darwin  mux tests: before=true  after=true    no-mux tests: false
  win32   mux tests: before=false after=true    no-mux tests: false

A second Windows problem mux: true does not fix

The control-dir safety block in open() is itself wrapped in process.platform !== 'win32' — so open() rejects a control-dir that is a symlink has no rejection to assert on Windows. Added the same if (process.platform === 'win32') return guard that the 0700 test three lines below already uses. Worth knowing that explicit mux alone would not have been sufficient.

Two sub-claims from my earlier triage, also corrected

• createSshProbeConnection already forces mux: false internally — that test needed nothing.
• The close() test I listed as already mux-explicit was a false grep hit on the string 'mux: master gone' in a stderr fixture. All eight failures share one cause, which is simpler than I reported.

Not included

ssh-connection.test.ts is not added to the Windows lane. This makes the suite capable of running there, but it's 40+ tests I can't execute from here, and over-including unverified files caused rounds #174 and #175. Adding it is a separate, checkable step — and given the lane is a required check since #176, the cost of being wrong is now everyone's PRs, not just mine.

---

Generated by Claude Code