Tests: collapse same-path cases in test_inbox.py (#497 step 5)

[renmengye]

Step 5 of #497 (fewer, behaviour-focused tests). Tests only; src/ is unchanged.

tests/test_inbox.py: 108 tests (106 passed, 2 skipped) → 90, 1.1 s → 0.7 s, with an identical covered-line set for outerloop.inbox (448/496, the same missing lines).

Every collapsed case was mutation-checked: the source was broken the way a regression would break it, a remaining test was confirmed to fail, and the source was reverted.

• Render kinds: note and base-moved share one branch in message_text.
• Check-log combos (8 → 6): the not-stale and dirty rows that only differ by sha string are merged. The not-stale-but-dirty case is kept: codex showed that a regression treating dirty as stale would pass without it.
• Kernel-instruction unfencing (4 → 1): reader never changes this render. Unfenced git text is pinned by the legacy base-moved test.
• Association "" vs None: these are identical after or "". None is kept.
• Malformed envelopes (6 → 3): one case per distinct decode rejection.
• Legacy mixed and legacy refusal: v2 is kept once for the envelope axis.
• Legacy panel wake ended: ended is refused in save_record, which test_runstate.py pins byte for byte.
• GitHub producer spoof values (2 → 1): the producer's source key is never read.
• The exact merge-command sentence in test_check_log... becomes a substring check. The full wording stays pinned by test_base_moved_advises_direct_submit.
• GLM collapsed the check-conclusion cases at first. The mutation check showed skipped is classified nowhere else, so those cases were restored.

Found while doing this, not changed here: no test pins kernel_text's "refusal:" prefix, and .coverage is not in .gitignore.

Built by GLM-5.3 (self-reviewed, coverage-gated, mutation-checked), reviewed by Codex (one dropped case restored in c6c9729), final read by Claude Opus 5.5.

[github-actions]

Round 1 — reviewed head c6c9729f — 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 findings: none.