Tests: fold repeated setup in test_syscall (#497 step 6)

[renmengye]

Step 6 of #497 (fewer, behaviour-focused tests). Tests only; src/ is unchanged.

tests/test_syscall.py was already close to one test per behaviour: no harness call-sequence pins and no large parametrized cross-products. This step only removes repetition. All 75 tests stay, the file goes from 1295 to 1219 lines, and the covered-line sets of outerloop.syscall and outerloop.syscall_cli are unchanged.

• About 30 per-test imports were moved to the top of the module.
• A _job() helper now builds the eval-launch job directories that five tests wrote by hand. Codex review found that the first version of the helper changed two fixtures: the train job lost its empty stderr file, and the big job gained an artifacts dir it never had. That moved _deliver_artifacts off its no-directory path. The helper now reproduces each original fixture exactly.
• write_req(typed=False) now writes the raw ABI files that two tests created by hand.

GLM also dropped three cases: [None] message items, whitespace-only withdraw reasons, and the symlinked sync-done marker. Codex named a realistic regression that each one is the only test for, so all three stay.

Left for a later step: split test_refresh_tool's eight defences; move test_parse_elapsed_reads_sacct_fields to test_compute.py.

Built by GLM-5.3 (self-reviewed, coverage-gated, mutation-checked), reviewed by Codex (two rounds; the three deletions dropped and the fixture drift fixed), final read by Claude Opus 5.5.

[github-actions]

Round 1 — reviewed head 60bcef12 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: no defects found.

No findings were reported by any review lens. Rejected findings: none.