Collect all case precheck failures without suppressing native execution and reports

[LunaStev]

Observed on PR #606 head 037baf248a9793be2b0563fb28972b5c4fb9714a (2026-09-12).

The Linux amd64 job runs an all-target source precheck under set -e before its native case runner. One source error stops the entire precheck, skips the native run, and leaves the unconditional report upload with no file.

Observed in the PR #606 Linux amd64 job:

1. Check every target case stops on the stale helper import in linux/amd64/test13.wave.
2. Run all Wave cases is skipped.
3. Upload reports No files were found with the provided path: wave-cases-linux-amd64.json. No artifacts will be uploaded.

The workflow uses:

set -euo pipefail
mapfile -t cases < <(python3 tools/case_manifest.py sources)
for source in "${cases[@]}"; do
  target/release/wavec check "tests/cases/$source"
done

Starting point: cases.yml.

Acceptance:

• Check every discovered case and aggregate nonzero results with source paths and diagnostics, then return failure if any check fails.
• Preserve a precheck summary/log artifact even when prechecking fails.
• Keep the native case run eligible once its compiler/std prerequisites succeeded, or isolate the all-target precheck in a separate job. Do not attempt runtime execution after compiler provisioning/build failed.
• Preserve the job/workflow failure status and distinguish precheck failure, runtime failure, and tests that never ran; do not manufacture passing case records.
• Verify with at least two failing sources and a later successful source so fail-fast behavior cannot silently return.

Related: #575 concerns native ARM64 diagnostic-step ordering; #576 concerns the contents of reports produced by the runner. This issue concerns the separate all-target precheck and the missing report caused before the native runner starts. Fixing the obsolete import alone does not correct this workflow behavior.