Warn when a repeated -t silently discards earlier targets

[Mohit-Ak]

-t is declared with nargs="+", so argparse registers it with the default store action. A second occurrence of the flag doesn't append to the first — it replaces it. -t pre_process -t simulation therefore parses to ['simulation'], and the generated batch script has one fewer step than the user wrote, with nothing printed to say so.

Confirmed against the real generated parser (cli/argparse_gen.generate_parser over MFC_CLI_SCHEMA, which is what args.parse uses):

invocation	parsed targets
-t simulation	['simulation']
-t pre_process simulation	['pre_process', 'simulation']
-t pre_process -t simulation	['simulation']
--targets pre_process --targets simulation	['simulation']
-t pre_process -t simulation -t post_process	['post_process']

On the two suggested fixes

The issue offered either a warning or action="extend". I went with the warning, but only after testing extend — it turns out to be actively wrong here, because targets has a non-empty list default:

invocation	current	action="extend"
-t simulation	['simulation']	['pre_process', 'simulation', 'post_process', 'simulation']
-t pre_process simulation	['pre_process', 'simulation']	['pre_process', 'simulation', 'post_process', 'pre_process', 'simulation']

extend appends to the default rather than replacing it, so -t simulation would silently start running all three targets — it converts a quiet drop into a quiet addition, which is worse: -t simulation is the restart-from-existing-data invocation, and the whole point of passing it is not to re-run pre_process. Every -t call site in the repo (test/case.py, test/convergence.py, bench.py) would be affected too. So: warning only, and parsing behaviour is left exactly as it is.

What the fix does

_warn_on_repeated_multi_value_flags counts occurrences of each list-valued flag in cli_argv and prints a yellow note when one appears more than once. It's driven off the CLI schema rather than a hardcoded -t: _multi_value_flags walks the command's own arguments plus everything it pulls in through include_common, and yields any argument declared nargs="+" or "*". That means --gpus and --only are covered by the same check — they have the identical shape and the identical failure mode — and any future nargs="+" option is covered automatically. Short/long/= spellings all count toward the same option.

End to end, on the exact command from the report:

$ ./mfc.sh run examples/1D_sodshocktube/case.py --dry-run -t pre_process -t simulation

 -t / --targets was given 2 times, but it takes a space-separated list and only the
 last occurrence is kept. Earlier values were discarded; pass them together as e.g.
 --targets A B.
 ...
                                    | --targets simulation

The banner underneath still reads --targets simulation, which is the behaviour the warning is there to explain.

Testing

toolchain/mfc/test_args_repeated_flags.py — 16 tests covering the repeated short form, the long form, mixed spellings, the = form, three repeats, targets reaching build via include_common, --gpus, and the schema walk itself. The silent controls matter as much as the warnings: -t pre_process simulation, -t simulation, and no flag at all must all stay quiet, and one test asserts parse() still returns argparse's unmodified result.

RED→GREEN was proved in a throwaway worktree with args.py restored from upstream/master and the tests copied in — using a behavioural test that imports only names present on upstream, so it runs identically on both sides:

RED  (unpatched args.py):  1 failed, 2 passed
     AssertionError: a target was silently discarded: nothing was printed to warn the user
GREEN (patched args.py):   3 passed

The 2 that pass on both sides are the controls (intended form, single target) — they confirm the failure is the dropped-target case specifically and not a broken harness.

Both CI gates, re-run against the committed SHA with a clean tree:

command	CI workflow	result
./mfc.sh lint	test-toolchain-compat.yml	705 passed, 9 subtests passed
./mfc.sh precheck -j 4	lint-toolchain.yml	7/7 checks pass

Fixes #1875

[sbryngelson]

doesn't use pr template