wait/job-output: report scheduler job-id rejection distinctly instead of UNKNOWN / tail -F hang

[ultimatile]

#50 made the scheduler layer classify an affirmative job-id rejection as JobIdRejectedError (a SchedulerError subtype), and status now reports it distinctly: an unknown id gets the Run not found report, and a known run whose recorded job id is rejected prints Job <job_id>: the scheduler rejected this job id (run <run_id> metadata may be stale or corrupt) with exit 1.

The other commands that query the scheduler with a recorded job id still conflate the deterministic rejection with transient states:

• wait and submit --wait: since status: run-id-shaped argument from the wrong project root raises SSHError instead of the Run not found report on Slurm #50, wait_for_job returns UNKNOWN immediately on the first poll for a rejection, so the wait loop terminates rather than hanging — but the CLI prints final state unknown (scheduler stopped returning data), a transient-sounding message that misdirects the user into retrying, and overwrites run.status with unknown.
• job-output --follow: the pre-tail status check swallows the rejection via its generic except SSHError and falls through to tail -F on an output path that can never exist, hanging indefinitely.

These paths are reachable only with stale or corrupt run metadata (recorded job ids come from scheduler acceptance, so a live submission never produces them), so #50 was deliberately scoped to status, the one command that also accepts raw user-supplied ids; this issue extends the distinct report to the remaining commands.

Proposal

1. Let wait / submit --wait distinguish a rejection: either surface JobIdRejectedError from wait_for_job to the CLI, or return a distinct terminal marker, and print the same rejected-id diagnosis status gives (exit non-zero, without clobbering run.status).
2. Make job-output --follow's pre-tail status check treat JobIdRejectedError as fatal instead of falling through to tail -F.
3. Reuse the recorded rejection fixtures in tests/rejection_fixtures.py for the new tests, mocking at the SSHManager layer.