Document the live config keys in the generated README

[florian6973]

Nine config keys are reachable only by reading the YAML. The generated README documents the five commands
and their artifacts, then says "any extra argument is a Hydra override" — true, and the reason nobody
discovers pipeline.

What to add

A Configuration section in README.md.jinja. Its content is roughly the live-key table from the audit
— the keys that select a branch and change what the commands actually do:

key	commands	non-default behaviour
do_resume	pretrain, supervised_train	reuse scratch checkpoints from a previous attempt, or discard them
work_dir	pretrain, supervised_train	relocate the scratch directory away from the output sibling
clean_work_dir	pretrain, supervised_train	keep the scratch tree instead of rmtreeing it
pipeline	preprocess_data	run an additional CLI (MEDS_transform-pipeline) before tensorization
pipeline_overrides	preprocess_data	Hydra overrides appended to that CLI
attach_labels	predict	emit predictions without boolean_value
seed	pretrain, supervised_train	null skips seed_everything entirely
splits	infer, predict	accepts null, a bare string, or a list
do_reshard	preprocess_data	false skips reshard_to_split, requiring split-sharded input
input_data_dir: null	predict	recover the workspace from the source artifact's manifest
do_overwrite	all	replace an existing artifact instead of refusing

The inert keys (max_seq_len, batch_size, num_workers, trainer.*, optimizer, model.*, the
output_* paths, the paths/ interpolations) need only a one-line pointer at configs/ — documenting
them individually documents Hydra, not this template.

Acceptance

• A Configuration section exists in the generated README covering every key above.
• The pipeline + do_reshard interaction is stated (see preprocess_data: do_reshard guard, split-table desync, and the untested default path #2) — it is the one combination a reader can
  otherwise only discover by hitting it.

Notes

One issue on purpose. Nine per-key issues would put nine agents in README.md.jinja and produce
nothing but conflicts.

Coordinate with #6, which adds determinism keys that belong in the same section.

From the Hydra configuration audit, §1, §2 and §6 (docs/CONFIG-AUDIT.md).

[florian6973]

Still needed, with two corrections after #9/#11 and the agent-docs change.

Drop the do_reshard row. It is no longer a config key — preprocess_data.py:127 passes
do_reshard=false unconditionally, documented at preprocess_data.py:115 as an invariant.

Drop the second acceptance bullet. "The pipeline + do_reshard interaction is stated (see #2)" — that
combination can no longer be expressed, and #2 is closed. The precondition it stood in for is already
documented: README.md.jinja has an "The input dataset must be sharded by split" section explaining the
layout, why meds-torch-data requires it, and the effect of a subject-filtering pipeline= on evaluation
coverage. What remains missing is the table of the other keys, which is the actual point of this issue.

One thing that has not changed: generated repos now ship a CLAUDE.md and docs/PORTING-A-MODEL.md, but
neither documents these keys. CLAUDE.md deliberately points at configs/ and says to configure through the
config tree rather than in Python; it does not enumerate what is there. So this issue is not covered by that
work, and the Configuration section still belongs in README.md.jinja.

Coordination with #6 still applies — the determinism keys it adds belong in the same table.

[florian6973]

Two more rows to drop, after #12: work_dir and clean_work_dir no longer exist. Scratch is always <output_dir>.work and is always removed once the artifact is published.

The do_resume row stays but its description changes — it now defaults to false, so the entry should say what turning it on does and when that is safe (continuing the same run), not "reuse or discard". The reasoning is already written up in the #: note above the key in pretrain.yaml, which the README section can draw on rather than restate.

That takes this issue from eleven keys to eight.