ci: eliminate hosted runner test flakes

[loadingalias]

Summary

• isolate RSA allocation evidence to the measured test thread
• avoid SIGPIPE while reading Linux CPU feature records
• apply the CPU probe correction to both RSA and AMX native gates
• remove unused test debug metadata from the cold RISC-V runtime lane
• give uncached RISC-V validation a truthful bounded timeout

Root causes

The PR head and squash-merged commit have identical Git trees. The first post-merge failures were nondeterministic test infrastructure: process-wide allocation counting observed unrelated harness allocations, while sed | head could terminate sed with SIGPIPE under pipefail.

The follow-up PR run proved both fixes, then timed out in the required RISC-V lane before tests began. Cargo Rail 0.25.0 has no riscv64gc-unknown-linux-gnu release archive, so that lane correctly remains cache-disabled and must support a complete cold all-features build.

Validation

• exact Cargo Nextest allocation test under full
• exact Cargo Nextest allocation test under std,full
• local RISC-V inner command shape with test debug metadata disabled: 1,084 library tests passed
• cargo clippy --locked --workspace --all-targets --no-default-features --features std,full -- -D warnings
• target catalog validation and exact RISC-V/POWER10 assertions
• just check-actions
• just check
• just push
• shell syntax and CPU fixture parser checks