Bring the code_aster adapter up to the level of code_saturne

[florian-simvia]

Where code_aster stands today

CodeAsterAdapter implements the run lifecycle and nothing else. Measured on
main with the _provides introspection the capability work added:

Domain	Method or declaration	code_saturne	code_aster
Status	read_progress	yes	—
Residuals	find_residuals_files, parse_live_residuals	yes	—
Probes	list_probe_files, list_profile_files, locate_probe_files	yes	—
Performance	find_performance_log, parse_performance, performance_columns	yes	—
Live control	apply_control, control_actions	yes	—
Restart	build_restart_args, read_restart_origin	yes	—
GUI	gui_argv	yes	—
Slurm	build_slurm_script, mpi_env	yes	—
Files	locate_case_file, find_run_config	yes	—
Doctor	doctor_checks	yes	—
Declarations	anomaly_file_names	4 entries	2 (base default)
Declarations	cleanup_log_names	6 entries	2 (base default)
Declarations	template_input_names	2 entries	empty

Derived capabilities, which is what the dashboard now shows:

• code_saturne: compare, control, gui, performance, probes, residuals, restart
• code_aster: compare

So a code_aster user gets four panels (Status, Compare, Log Tail, Recent
Errors) and three buttons (Run, Kill, Clean).

Parity is not "implement all eighteen"

Some of these are meaningless for code_aster and should stay unimplemented on
purpose. A linear static analysis has no residual curve to plot, and the
capability mechanism already handles that correctly: an unimplemented method
means the panel simply does not appear.

The first task of this issue is therefore a triage, not code. For each row
above, decide: meaningful for code_aster, meaningless, or meaningful but out of
scope for now. That triage needs someone who knows code_aster; the rest is
mechanical once it is settled.

Blocker: DIAGNOSTIC JOB is never found in the shipped example

Independent of parity, and worth fixing first because it makes the shipped
example misreport every run.

_ensure_mess_entry injects F mess RESU/LOGS/run_solver.log R 6 into the
.export only when the file declares no F mess line at all
(code_aster.py:76). The shipped examples/codeaster-cube/TEMPLATE/study.export
declares one on line 18:

F mess output.mess R 6

So the injection is skipped, code_aster writes its message file to
output.mess, and detect_outcome looks at
RESU/LOGS/run_solver.log (code_aster.py:168), which never exists. It returns
None for ever, and the case is finalised by the PID-death fallback as
FAILED — even after a perfect run.

The integration test does not catch it: tests/integration/test_code_aster_adapter.py
builds an .export with no F mess line, so it exercises the branch where
the injection happens. It is also gated behind CSAUTO_CODE_ASTER_DOCKER_TESTS=1
and never runs in CI.

Two candidate fixes, both needing a code_aster opinion: rewrite the existing
F mess line rather than leaving it, or have detect_outcome read the path the
.export actually declares instead of a hardcoded one. The second is cleaner
and also handles hand-written templates.

Three more design questions in the current adapter

Each needs a decision from someone who knows the solver:

1. shared_dir_names = ("MESH", "RESU") while results_dirname = "RESU".
   The two meanings collide: shared dirs are template-level directories
   symlinked into the runs dir and bind-mounted read-write into every
   container, while results_dirname is per-case. As written, if a RESU
   directory sits beside the template, every case would write its results into
   the same shared directory. csauto prepare already warns
   source directory missing for copy RESU: RESU on the shipped example.

2. run_argv returns [] and build_run_command is overridden wholesale,
   so code_aster bypasses the generic runtime wrapping in execution.py that
   code_saturne and the stub use. That may be unavoidable given run_aster's
   shape, but it means container mounts, labels and cleanup are reimplemented
   rather than shared, and the two implementations can drift.

3. Shell quoting in build_run_command. The docker and singularity commands
   are assembled as a bash -c string mixing quoted fragments (bind_links,
   solver_cmd) and unquoted ones (container_case, the --label value). A
   runs directory containing a space breaks the command.

Suggested order

1. The F mess blocker, plus an integration test that uses an .export with an
   existing F mess line, so the shipped example's shape is covered.
2. The triage above, recorded in this issue as a checklist.
3. anomaly_file_names and cleanup_log_names, which today miss
   RESU/LOGS/run_solver.log. Cheap, and they immediately improve the Recent
   Errors and Clean panels.
4. read_progress, so the status table shows advancement instead of an empty
   column.
5. Whatever the triage keeps, in descending user value. performance_columns
   and apply_control are likely the next two.
6. The three design questions, each as its own PR.

Definition of done

• Every capability code_aster can meaningfully feed is implemented, and every
  one it cannot is documented in this issue as a deliberate omission.
• csauto doctor on a code_aster campaign reports the panels and capabilities
  that triage decided on.
• examples/codeaster-cube runs to completion and is reported DONE, verified
  with the docker image rather than by reading the code.
• The integration test covers the shipped example's .export shape and runs in
  CI, or the reason it cannot is written down.