Test audit 3: finish the test policy cut (the owner's five answers)

[tamashi095]

Test audit 3: finish the test policy cut (the owner's five answers)

Closes #1726

The second owner-approved policy cut of the 2026-10 test audit. It applies the owner's answers (2026-10-19) to the five questions that test audit 1 (#1722) left open. The spec lists every change by answer.

What changed

1. The 3 scalar-oracle pairing tests that were the only readers of the initial-matrix trace are deleted, together with that trace and the scalar state trace that only they read (about -1,100 lines). The oracle's render arithmetic is unchanged.
2. The math EXP2/LOG2 digest is kept. It is the native witness of the open AArch64 defect Make the compressor's AArch64 release build bit-identical to the browser (LANE-3) #1019, and the PR that fixes Make the compressor's AArch64 release build bit-identical to the browser (LANE-3) #1019 deletes it.
3. The host-core limiter_linked_session pin, which claimed to be a second cross-target corpus owner, is deleted. Its case moves into the G5 corpus, the single owner: the new case maximum/link_unlink_relink covers the limiter's link decision. G5's 5 earlier limiter cases all ran unlinked. The new pin is the same at Scalar, Simd4 and Simd8, and in the wasm leg (153 cases, 0 mismatches). The session checks of the old test stay.
4. gain_pan_profile phase_profile is deleted, together with the graph and rack phase-profile probes and a tool example that used them. No product code used them. Rack's test-support feature is now empty, so it is removed.
5. One-line size_of counts inside larger tests are trimmed. Where a line guarded a product relation, it is now that relation, and each one is red on its mutant. A new test-support accessor, graph::test_only_carry_counts, gives successor_swap exact carry-table counts. capi resource_lifecycle is unchanged: its 3 size_of lines feed exact caps.

21 files; -2,808 / +631 lines. No product code changed.

Review

An adversarial verifier gave READY, with 3 minors and 2 nits. All are folded in.

Evidence

On the tree merged with main fdb1c44, all of these pass: fmt, clippy -D warnings --all-features on 9 packages, 1,207 tests, check-workspace-policy.sh, run-wasm-gates.sh --without-v8-spill and both test-support CI scripts. After the fold, 675 tests passed again.

Residual risk

The new G5 case runs the linked limiter body on NEON under G5 for the first time, and nobody has checked its pin on AArch64 yet. If the AArch64 release leg goes red, the cause is an arm64 digest difference that the new case exposes, not a regression this PR makes.

CI folds

• The first CI run failed in the lint job on two private intra-doc links in crates/true-peak-limiter/src/corpus.rs. ea1273a makes them plain code spans; all 31 steps of that job then passed locally.
• The second run's workspace debug test job hit its 15-minute timeout while still passing tests (recent main runs take 14.5-14.7 minutes). One rerun of the failed job passed in 10.5 minutes. The timeout margin is tracked as its own issue.

🤖 Generated with Claude Code