Investigate NVHPC nvfortran serial smoke validation

[jaharris87]

Deferred from PR #254.

Scope: Investigate a conservative NVIDIA HPC SDK nvfortran serial smoke/install-consumer CI path on GitHub-hosted Linux runners.

Why deferred: The issue #245 trial installed NVIDIA HPC SDK 26.3 (nvhpc-26-3) and built fTimer successfully, but the smoke/install-consumer executables aborted at runtime with DEALLOCATE: memory at (nil) not allocated. The first pass removed the NVHPC job and avoided making a support claim.

Acceptance criteria:

• Reproduce or rule out the hosted-runner DEALLOCATE: memory at (nil) not allocated abort with a minimal serial smoke path.
• If reliable, add a conservative nvfortran CI job that configures, builds, runs basic smoke/example coverage, and verifies installed-package consumer behavior.
• Keep NVHPC MPI, NVHPC pFUnit, and NVHPC OpenMP out of scope unless they turn out to be cheap and reliable.
• Update README/design support claims only for paths that pass CI.

Risk / impact: Low for current releases because NVHPC is not currently claimed as validated, but support remains unknown until the runtime abort is understood.

[jaharris87]

Investigation update from PR #258:

We should treat the hosted-runner nvfortran serial smoke path as not addressable right now. PR #258 tried to make NVHPC work on GitHub-hosted Ubuntu, including NVIDIA HPC SDK 26.3 (nvhpc-26-3) and a 25.11 comparison. In both cases fTimer configured and built successfully with FC=nvfortran, but the runtime smoke/install-consumer executables still aborted on the hosted runner.

The failure pattern was consistent: ftimer_phase0_smoke, both installed-package consumers, basic_usage, and the enabled instrumentation facade executable aborted around the first real fTimer start path, with NVHPC/libc traceback output ending in __libc_free / invalid free or segfault-style cleanup. The disabled instrumentation facade executable and install-only checks still passed, which kept pointing at fTimer-linked runtime execution rather than package installation alone.

A diagnostic run with FTIMER_START_TRACE=1 could make the direct ftimer_phase0_smoke executable pass through start, stop, and finalize, but the normal untraced CTest run still failed. Attempts to stabilize the path with conservative Fortran-side changes and NVHPC compile options, including -Mrecursive, did not produce a reliable normal smoke/install-consumer job. Some experiments moved the failure boundary or made diagnostics pass, but none produced a clean CI-validatable path without carrying intrusive diagnostic behavior.

I am closing/abandoning PR #258 rather than merging partial diagnostics or broadening support claims without CI validation. A future attempt is probably only worthwhile with access to a local machine/container where NVHPC can be run interactively under a debugger or allocator tooling; iterating solely through GitHub Actions was too slow and opaque for this runtime/compiler cleanup failure.

[jaharris87]

Mapped into the v1.0 readiness umbrella #348 as V1-NB-04: NVHPC serial smoke validation.

Disposition for v1.0 planning: this remains non-blocking because current README/release evidence do not claim NVHPC as a validated compiler path. v1.0 should keep NVHPC in the plausible-but-unvalidated bucket unless this issue is resolved with reliable smoke/install-consumer evidence and the support matrix is updated accordingly.