Delete the batch 2 redundant tests (test audit TA5, #1734)

[tamashi095]

Test audit phase B, PR 5 (TA5). This is the last PR of the audit. It deletes the tests of the batch 2 sample (200 random KEEP tests) that are redundant both by measurement and on reading.

Closes #1734. Adds the spec for #1733. Refers to #1728 and does not close it.

Summary

• 97 tests deleted. Batch 2 measured 162 of the 200 sampled tests as redundant. Each of the 162 was then read against the AGENTS.md test-value question. 97 are deleted, and each deleted test's answer names the remaining test that defends its claim. 65 are kept, each with a one-line reason:
  • the acked-batch question;
  • the only red test of a MUTATIONS.md row;
  • known AArch64 defects;
  • ABI or wire layouts;
  • allocation, count or ceiling claims;
  • exact edges;
  • randomized differentials;
  • reproducers.
• 31 helpers deleted (465 lines). Each went only after clippy reported it unused, and no #[allow] was added. 2 test files are deleted.
• MUTATIONS.md: rows that named a deleted test now name the remaining catcher. The "TA4 deletions" ledgers that TA4 appended are removed. The list lives in git history and in the comments on Delete the tests that mutation measurement found redundant (test audit TA4) #1731.
• Give the workspace debug test job a real time margin #1728: the test-debug-a timeout goes from 15 to 22 minutes, which is 1.5 times the slowest recent green run (14:41). It still took 14m29s after TA4.
• 12 survivors. 3 are equivalent, 1 unreachable and 2 dead code. 6 are real gaps, which go to Close the batch 2 mutation-survivor gaps #1733 with no new tests in this PR. The most important gap: bench-support record_peak_window could be a no-op and the Shrink the peak heap of a live-only commit #1615 heap ceilings would still pass.

The test-by-test list with each answer and its catchers is in the comments on #1734.

Finding for any later cut

About 12 random mutants per test overstate redundancy. 81% of the sample measured redundant, but only 49% was safe to delete after reading. B2's pool-wide projection does not stand. The spec's last section states what a reopening brief must check first.

Gates (author, final tree)

• All 31 steps of the qualification lint job: PASS.
• cargo test --locked --all-features on the touched packages: PASS (2,276 tests after the review fold; parametric-eq 122 after the last restore).
• A fresh verifier ran two rounds. Round 1 restored 5 tests and kept 28 more. Round 2 restored 1 reproducer. Final verdict: READY.
• Root removed the TA4 ledgers and numbered the specs. check-conformance-boundaries.sh and the spec lint scripts PASS.

-5.2k / +0.3k lines over 88 files.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}