MinGram across the grid, and the bnd_wpd_extcaps sweep

[cimeister]

Three commits, cut from the current tip so this fast-forwards.

MinGram now covers every language. The grid had MinGram for English plain and bnd_wpd only, so make_intrinsic_table.py dropped the trainer entirely, since a trainer earns a column group only when it covers the same languages as the leading one. 18 new cells complete plain, bnd_w, bnd_wp and bnd_wpd for en, de, fi, ru and ar. table_intrinsic.tex goes from 46 to 56 of 72 cells, and the two trainers can now be read against each other. The scheme ordering holds under both: MinGram eval means of $-9.85$, $-2.70$, $-0.96$ against BPE's $-9.30$, $-2.18$, $-0.46$.

Every corpus was already cached, so only the trainer ran: 1.48 node-hours for the 16 non-English cells.

bnd_wpd_extcaps measures as no different from bnd_wpd_caps downstream. Three seeds, every setting matched to the committed BPE runs, paired by seed:

comparison	mean	SE	t(2)
vs plain	+0.006026	0.000128	47.05
vs bnd_wpd	+0.000709	0.000268	2.64
vs bnd_wpd_caps	+0.000247	0.000233	1.06

The difference against bnd_wpd_caps is smaller than either arm's seed standard deviation, and compression agrees (+0.75% against +0.72% on Goldfish). The commit adding the arm said the probe corpus could not answer whether the shared word entry is reused; at this scale it makes no measurable difference either way.

Two things you should look at rather than take on trust.

fineweb_ru_5gb_plain_mingram_v34685 is vocabulary 34,684, one short, and is committed as it is. The cause is in the MinGram trainer: the target check at script_bpe/tokenizers/mingram/trainer.py:57-64 sits outside the inner EM loop, so once the size reaches the target the remaining sub-steps still prune. Its log shows EM 2.1. Model size 34,685 then EM 2.2. Model size 34,684. Any MinGram cell can overshoot by up to num_em_iterations - 1; the rest landed exactly by luck. merge_manifests.py correctly refuses to rewrite manifest.json for this grid, and I left that refusal in place rather than passing --no-check, so manifest.json is untouched here.

The Goldfish measurement was run twice and the first was thrown away. Run as python marker_experiments/downstream/eval_goldfish.py, sys.path[0] is the script's directory, which holds no marker_experiments package, so the import resolved through the editable install to a different checkout on a different branch. It measured 10 cells against that branch's pretokenizer before raising KeyError: 'ExtCapsBoundaryScriptPretokenizer' on the eleventh. Those 10 were already written and would have looked ordinary in review. They were discarded and the measurement rerun with PYTHONPATH pinned. Worth knowing if you ever run that script from a worktree.

submit_extcaps.sh also picks up four fixes already made in the MinGram launcher: pinned SMOKE and TAG_SUFFIX, a clean-tree guard that uses git status --porcelain rather than git diff so it sees untracked files, a squeue failure that aborts instead of yielding an empty in-flight list, and a required ACCOUNT.