Tests: slim test_attempt_review.py to behaviour-focused cases (#497 step 2)

[renmengye]

Step 2 of #497 (fewer, behaviour-focused tests). Tests only; src/ is unchanged.

tests/test_attempt_review.py:

	before	after
tests	122	81
covered lines of outerloop.attempt / outerloop.orchestrator	901 / 497	901 / 497, the same line sets
runtime	about 43 s	about 45 s (the slow scenarios all remain; the deleted cases were cheap)

• Parametrized cross-products collapse to one case per distinct path. test_review_submit_parks_and_delivers_verdict goes from 12 cases to 4 (plus one gpus=0), test_publish_review_fast_forwards_and_applies_floor from 24 to 6 (plus the interrupted report-only recovery), and smaller merges are made in the reply, preflight, ending and scope tests.
• commit() and the review-GitHub fake are extracted as helpers, and dead contract constants are dropped. One test that byte-duplicated another is deleted.
• New assertion: the floor test carried an expected value it never used, and its outcome (the "ledger row is unchanged" floor note) was asserted nowhere in the suite. It is now.
• Codex review found three collapsed axes that do change an outcome: a non-main base for _submit_preflight, gpus=0 (CPU submits charge nothing), and interrupted report-only publication recovery. One case each is restored (7929b87).

Built by GLM-5.3 (self-reviewed, coverage-gated), reviewed by Codex, final read by Claude Opus 5.5.

[github-actions]

Round 1 — reviewed head 845ede2a — 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 submitted by any review lens. Rejected: none.