Fix time check

[xylar]

Partially addresses #12 but with an open question.

What was wrong

The time checks inferred a run's length from its endpoints and its cadence from its first interval, so nothing looked at what happened in between. A ctrl file whose time axis holds 2015, 2016, 2299, 2300 starts in the right year, ends in the right year, and has a 365-day first interval — it passed every time check while being 282 time steps short of the 286 required. Decimated axes, mid-run gaps and dropped segments all passed the same way. I confirmed this against main rather than inferring it: the old checker reports 0 errors on that file, the new one reports 2.

The snapshot (x,y,z,t, i.e. litemp) check had the matching gap. It validated only the nominal years a file happened to contain, and built its set of valid years partly out of the file's own last time step, so whatever a file ended on was correct by construction and nothing ever looked for snapshots that should have been present and were not. A litemp file holding a single snapshot at an arbitrary year reported 0 time errors on main; it now reports 1.

What changed

Reconstruct the expected axis and compare. The nominal years come from experiments_ismip7.csv and are encoded with the variable's ST/FL convention, giving the exact list of timestamps the file should hold. One comparison then subsumes the step count, the cadence, the first and last timestamps and the duration — and it can say which years are missing rather than only that something is off. Messages collapse consecutive runs, so a 99-year hole reads as 2101-2199 rather than as ninety-nine separate findings.

Filename year range moves into the time check. _check_naming now only parses the range and checks it is well formed, handing the parsed years to _check_time, which decodes the file anyway. Previously both functions decoded the dataset separately and each carried its own copy of the ST/FL year offset. Errors about the year range therefore now count under Time rather than Naming in the summary — expected, and the main reason the golden log moved.

Required snapshot sets for x,y,z,t. The required set is the run's final year, the century marks falling inside the run, and — for historical only — the run's first year. historical is the one experiment whose start year the modeller chooses, so it is the one whose initial state is not recorded anywhere else; a projection's initial state is the historical run's final state, already reported as historical's last-year snapshot. Missing snapshots are now errors.

A file is checked as far as it can be checked. _run_variable_checks used to abandon a file the moment _check_naming reported anything, so one unrecognised ESM name meant the numerical, spatial, time and attribute checks never ran — the report read "1 error(s)", the modeller fixed the name, re-ran, and only then learned about everything else. That also hid exactly the time errors this PR adds. Only a missing x/y dimension now stops the rest. An unrecognised region costs just the three checks that index by it (value range, grid extent/resolution, crs), each of which now says it is skipping and why; previously _check_spatial quietly fell back to the GrIS extents and checked the file against the wrong ice sheet's criteria.

One decision that needs @hgoelzer's confirmation

The repo documents two different sets of required snapshot years, differing only over 2000. ISMIP7_variable_request.csv lists "(1900 if in historical), initial state, 2014, 2100, 2200, and 2300"; the README, checker.py and generate.py all instead used {1900, 2000, 2100, 2200, 2300}. Since the check now errors on missing snapshots, the required set has to be right in a way it did not before.

This PR follows the data request and does not require 2000, but tolerates it when present — the generator no longer writes it, and a file that carries it is not failed for a year still in dispute. That keeps files written to the earlier README passing. See #12.

What reviewers will notice

The golden log shrinks by 81 lines: four verbose time lines per file collapse into one, e.g. Time step: 365 days / Experiment lasts 2 years / First time stamp correctly set to… / Last time stamp correctly set to… become Time axis: 2 annual FL time step(s) covering nominal years 2013-2014, as experiment 'historical' requires: OK. Error counts on the reference dataset are unchanged at zero.

Error counts on broken files will rise, sometimes sharply, because files that used to abort after their naming block now run the remaining four blocks. Those errors were always present; the checker was not reaching them.

TIME_STEP_MIN_DAYS and TIME_STEP_MAX_DAYS are gone — the 365-vs-366 latitude they encoded comes free from comparing real calendar dates. generate.py imports CENTURY_SNAPSHOT_YEARS from checker.py rather than keeping a second copy, so the generator and the checker cannot drift apart again.

Tests

51 tests, up from 39. New coverage for the axes that used to pass: hollowed out between correct endpoints, decimated, shifted by a year, duplicated, disagreeing with the file name, and written with the wrong ST/FL convention; plus litemp missing a required snapshot, carrying an unrequested year, and carrying the tolerated 2000. A full-length 286-year ctrl fixture is added — nothing exercised the real ISMIP7 projection length before. Two tests assert the decoupling directly: a file with a bad ESM name still reports its time errors, and a file with an unrecognised region still gets everything that does not depend on the region.

Known gap

litemp is non-mandatory and the golden-log fixture does not pass include_non_mandatory, so the snapshot path has no golden-log coverage. The dedicated tests cover it. Adding it to the fixture would be worth doing but would add several hundred lines to the reference file, so I left it out of this PR.