docs: explain LAMMPS exploration extra files - #393

Thanks for finishing this bookkeeping - Closes #304 is correct. The feature landed in eb779cb (#306, 2025-08-29), six days after #304 was filed, but #306's body said "fix issue #305" (an unrelated resubmit crash), so #304 never auto-closed. #306 also claimed "The docs were updated too" while git show eb779cb -- docs/ is empty, so the remaining gap really is prose. I traced the mechanism end to end and it works: npt_task_group.py stores the basename and adds the file to every task, prep_lmp.py writes every key of files() into the task dir, and run_lmp.py symlinks everything under the task path into the LAMMPS work dir before the freeze step. The docs are accurate.

One thing to fix, with two surfaces.

The new test's fixture is degenerate in exactly the two dimensions the new paragraph makes claims about. The paragraph asserts (a) the file is copied "into every exploration task directory" and (b) it "is available there by its basename". The single new test pins neither, for two independent reasons, so one fixture change closes both. Details are in the two inline comments.

Evidence for (b), since it is the less obvious one - I reverted just the basename call in dpgen2/exploration/task/npt_task_group.py:

-        self.input_extra_files = [Path(ii).name for ii in input_extra_files]
+        self.input_extra_files = [str(ii) for ii in input_extra_files]

All 5 tests in tests/exploration/test_make_task_group_from_config.py still pass. For contrast, deleting the add_file loop that the test does cover fails it with KeyError: 'SiC_ZBL.txt', so the test is not vacuous - it just stops one step short of the property the docs advertise.

A suggested fixture that closes both at once: write the file into a subdirectory, pass sub/SiC_ZBL.txt (so basename stripping is observable), set "Ts": [100, 200] (so more than one task exists), and assert over every task in the group rather than tgroup[0].

Non-blocking notes, recorded so nobody has to rediscover them:

• The branch no longer merges. git merge-tree --write-tree upstream/master 23dee80 conflicts in docs/input.md, because #372 added a #### PLUMED CV candidate filtering subsection into the same gap. On rebase the new paragraph should stay immediately after the n_sample paragraph, before that #### heading - appending it afterwards would make it read as part of the PLUMED subsection.
• input_extra_files also exists on customized-lmp-template, where it means something different (files for the shell commands) and is genuinely not propagated: the inner lmp_tgroup.set_lmp(...) in _make_customized_task_group is called without it, so those files never reach the generated LAMMPS tasks. Your paragraph anchors to task_group[lmp-md] so it is not wrong, but that is a real separate bug worth its own issue.
• The lmp-template half of #306 is still uncovered: deleting its add_file loop leaves all 62 tests under tests/exploration/ green. Out of scope here, just noting it did not come with #306 either.

Two things I checked and am explicitly not raising, so they do not get re-opened: assets/dp/SiC_ZBL.txt not existing is fine - the whole example block uses illustrative paths (/file/prefix, relpath/to/confs/*, template.lammps); and writing the fixture into the CWD matches how every test under tests/exploration/ has worked since 2022.