fix: uniform level gating + step-count/UTC/-i ordering bugs in ingest-model.sh

[mrauhala]

Summary

Four bugs from a focused ingest-model.sh audit. The first two are the gate inconsistency cluster; the last two are small high-confidence correctness fixes that ride along.

F1 — uniform level gating

The SFC/PL/ML blocks had three different gates — only ML checked its input pattern existed. icon.cnf (area=europe) defines MODEL_RAW_SFC and MODEL_RAW_ML but no MODEL_RAW_PL, so every icon-europe run entered the PL block with an empty pattern and fired gribtoqd at a bare directory, failing every cycle. Now every completeness check and run block is guarded by [ -n MODEL_RAW_<lvl> ], matching ML.

F2 — drop eval from PL/ML step counts

eval gribstepcount "$INFILE_PL" expands the glob into separate args; gribstepcount's local FILES=$1 then counts only the first matched file. The SFC form (no eval) passes the glob as one string and lets grib_get $FILES expand it over all files. INFILE_* are already fully resolved upstream, so the eval was both unnecessary and wrong. All three checks now use the SFC form.

F3 — -t now parses as UTC

-t used date +%s (local) while -i and all RT_DATE_* use date -u. On a non-UTC host (e.g. FMI/Helsinki EET) this shifted the requested reference time by the TZ offset. Added -u.

F4 — -i <file> ordering

-i without -t ran the latest() incoming-dir scan first and could exit "No data available" before reading the -i file — exactly the case where incoming may be empty. And even on success, RT_DATE_*/OUTNAME/TMP were derived from the latest() time, then RT was overwritten by the file's time afterwards, so the output filename used the wrong reference time. Reordered: -i determines RT up front, before the derivations; latest() is skipped when -i is given.

Notes

• No behaviour change for the common models (all define SFC+PL, single combined input file per level), so the matching/completeness logic is identical there.
• Validated with bash -n and shellcheck (no errors). I can't exercise the grib_get/qdtools pipeline on macOS — the real check is a two-runs-no-change test on a host plus an icon-europe run showing no PL gribtoqd attempt.
• Touches the shipped script, so it changes RPM content: needs a version bump before merging to master (else the release workflow rebuilds/clobbers v26.5.22). Left out so you pick the version.

Test plan

• CI builds RPM on Rocky 8/9/10
• icon area=europe: run logs no PL conversion attempt (PL gated off; no gribtoqd on a directory)
• A model with multi-file PL split by step range: …_pressure.sqd completeness now matches on the second run (no perpetual re-convert)
• -t 20260522T12 on a non-UTC host yields reference time 12:00 UTC
• -i /tmp/file.grib with an empty incoming dir converts the file instead of exiting "No data available"