Address leftover review comments from merged PRs #2083

Follow-ups to review comments left open on PRs that are already merged. Each item was checked against current master before it was changed. Results do not change: the Fortran edits are comments plus one explicit real(wp) cast, and index.html is whitespace only (git diff -w is empty).

Bug

• ci: identify faulted node via sacct when the .out marker is missing #1835, submit-slurm-job.sh (thread): the sacct fallback that finds the faulted node had no || true. Under set -euo pipefail, a failing sacct ended the script without a message on the node-recovery path. It now ends in || true, like the other sacct call (~L198).

Minor

• Allow a GRCBC inflow to start up smoothly #1850, inflow ramp (thread): f(t0) = f0 + (1 − f0)(1 + tanh(−3))/2, which is f0 only to within about 0.25%·(1 − f0). case.md now says "≈ f0", and the comment in m_derived_types.fpp says the values at t0 and t0 + τ are approximate.
• Require the full inflow state when grcbc_in is enabled #1854, case_validator.py (thread): the grcbc_in inflow fields are built as one required list, and missing is computed from it in a single pass. The fields and the error message order are unchanged.
• Choose ib_neighborhood_radius from the thinnest rank, not the widest #1866, examples/3D_ibm_neighborhood_radius/case.py (thread): SUMMARY no longer hard-codes 16 x 2 x 2. It reads TOPOLOGY (default 16,2,2), computes the rank count and the extent in each direction, and uses the thinnest rank in any direction. With the default topology the output is byte-identical.
• Add the thin-plate force case that measures issue 1849 #1868, examples/2D_ibm_thin_plate_force/case.py (thread): removed the unused LEVELS. The comment above it referred to "L4", so it now says NCELL=16.
• Extract m_eos and add a single EOS family registry #1870, case_validator.py (thread): value or 0.0 is replaced by an explicit is None check.
• Probe the node with a solver binary, not just syscheck #1871, preflight.sh (thread): the "Illegal instruction" text match is kept on purpose. A comment now explains why: some launchers do not pass exit code 132 on.
• Stop amdflang GPU builds relinking the whole device image every time #1904, CMakeLists.txt (thread 1, thread 2): CMAKE_LINK_DEPENDS_USE_LINKER OFF now applies only when the compiler is LLVMFlang AND MFC_OpenMP. That combination is the amdflang offload build (MFCTargets.cmake adds --offload-arch under it), so GPU builds behave as before and LLVMFlang CPU builds keep the linker's dependency tracking.
• docs: add five flapping-flight and jet runs to the showcase #1912, docs/index.html (thread 1, thread 2): all 40 sims rows are now column-aligned to one width per field. The old rows did not share columns, and the new names are too long to fit the old ones.
• test: compile thermochem kernels with the configured Fortran compiler #1943, test_thermochem.py (thread 1, thread 2): the one-line "Cray" flags dict is now split across lines. The cached compiler probe moved to _probe_compiler(fc), so the cache is keyed on $FC and a changed $FC is probed again.
• Label time_data.dat and io_time_data.dat columns by what they measure #1949, m_start_up.fpp (thread): the comments now say time_final is the max over ranks of each rank's fastest RK stage, and io_time_final is the max over ranks of each rank's mean s_save_data time.
• Fixed-dt runs: keep the clock on t_step*dt so restarts continue bitwise #1951, m_start_up.fpp / m_time_steppers.fpp (thread 1, thread 2): removed the "1290 ulps by step 27000" figure and kept the reason. The fixed-dt clock now uses real(t_step + 1, wp)*dt and (real(t_step, wp) + 0.5_wp)*dt, in both s_perform_time_step and the IB stage time, so the two stay the same expression. The integer-to-real conversion is exact either way, so the values are bitwise unchanged.

Not changed

• docs: add five flapping-flight and jet runs to the showcase #1912, "512 GCDs" (thread): the PR first had "240-512 GCDs". A later commit on that PR (f6f6ce1, "Update accelerators for Seagull simulation") deliberately set it to "512 GCDs", and the merged PR description matches. There is no case file in the repo to check it against. Going back to the range would undo that deliberate edit, so the entry stays as it is.

Verification

• ./mfc.sh format: clean
• ./mfc.sh precheck (formatting, spelling, toolchain and source lint, doc refs, param docs, example validation): all 7 checks pass
• ./mfc.sh lint (ruff + toolchain unit tests): 794 passed, 2 skipped
• pytest toolchain/mfc/test_thermochem.py with CCE ftn: 16 passed, 1 skipped
• bash -n on preflight.sh and submit-slurm-job.sh: OK
• ./mfc.sh validate on both edited examples: pass. A grcbc_in case with fields removed still reports them in the same order.
• ./mfc.sh build -j 8 (Frontier, CPU): OK
• ./mfc.sh test -j 8 -% 10 on a Frontier CPU node: 72 passed, 0 failed
• ./mfc.sh test -j 8 -o IBM (includes the fixed-dt prescribed-kinematics and moving-IB goldens): 61 passed, 0 failed
• ./mfc.sh test -o <6 grcbc UUIDs> (1D/2D/3D grcbc x/y/z): 6 passed, 0 failed

Status: a fix/implementation for this is already written, in a pull request that has been closed for queue management.

• Closed PR: Address leftover review comments from merged PRs #1959
• Branch: ai-review-followups on the sbryngelson fork, at 91422f0389 (branch retained, not deleted)
• Size: +80/-68 across 12 files, 10 commits
• Where it sat: MERGEABLE; CI 19 checks queued (master was merged in today, so CI had not reported yet)
• Reopen with: gh pr reopen 1959 -R MFlowCode/MFC

It was merge-clean and up to date with master, and it is small.

Tracked with pr-to-do. Filed with Claude Code.