mado and mdl disagree on 36 of the 52 acceptance fixtures

[akiomik]

The acceptance comparison is not clean, and nothing recorded that. Raised while
fixing #396, which repaired the setup script that prepares this corpus.

What the comparison shows

scripts/acceptance/test.sh runs mdl and mado check --output-format=mdl
over the same fixtures, each under one fixed config (scripts/acceptance/.mdlrc
and the repository's mado.toml), and writes tmp/mdl.txt and tmp/mado.txt.
Diffing those two files on a corpus prepared by setup.sh:

• 36 of 52 fixtures differ.
• 103 findings mado reports and mdl does not.
• 30 findings mdl reports and mado does not.

Both files use the same path:line: MDxxx text form and absolute paths, so
these are real disagreements, not formatting.

mado only		mdl only
MD007	40	MD031	20
MD041	31	MD007	3
MD003	6	MD032	2
MD025	5	MD022	2
MD031	4	MD028	1
MD022	3	MD024	1
MD040, MD038, MD020, MD006, MD004, MD002	2 each	MD013	1
MD028, MD014	1 each

MD007, MD041 and MD031 are all listed in .mdlrc, so this is not a
question of one tool having a rule the other was not asked for. MD031 is the
interesting one: it appears on both sides, so the two tools are finding it
in different places rather than one of them missing it.

Why it is worth a look now

The corpus these numbers come from is what setup.sh is supposed to produce:
the fixtures upstream pairs with an X_style.rb are set aside, because those
are written against a rule set of their own that test.sh cannot apply. So
this is the intended comparison, and it is the one that disagrees.

Two things to separate before reading much into the totals:

• Whether .mdlrc and mado.toml really do ask for the same thing. They are
  written in different formats and were not derived from one another.
• Which differences are mado bugs, which are mdl bugs, and which are documented
  intentional divergences. Nothing in the repository currently says.

Note on how this went unnoticed

setup.sh's exclusion loop has never run in the committed script — see #396 —
so anyone cloning fresh compares all 88 fixtures instead of these 52, on a
corpus that includes documents designed for other rule sets. The numbers above
come from a checkout where the exclusion had been applied by hand before the
script was committed.

test.sh also has no diff and no assertion: it writes two files and stops.
So there has never been anything to notice, which is #400's subject from a
different angle.

[akiomik]

Found while reviewing #396, and it may account for a large share of these
numbers: the two tools are not configured alike. Before reading the
divergences as mado bugs, this has to be ruled out.

test.sh runs mdl with scripts/acceptance/.mdlrc, which lists rules and sets
no parameters, so every rule runs at mdl's own defaults. It runs mado with no
--config at all, so Config::resolve loads whatever mado.toml the working
directory yields — at the repository root, the tracked one, which sets a dozen
rule parameters.

Where they differ:

	mdl default	repo mado.toml
MD007 indent	3 (rules.rb:142)	4
MD013 code_blocks	true	false
MD013 tables	true	false
MD029 style	one_or_ordered	one
MD046 style	consistent	fenced
MD024 allow_different_nesting	false	true

MD007 is the single largest divergence in the run above — 40 findings mado
reports and mdl does not — and mado is checking it at indent = 4 against
fixtures mdl is checking at 3. MD013's two flags are inverted between them as
well.

Two consequences:

• The totals in this issue cannot be read as compatibility gaps until the
  configs are matched. Some part of them is just the two tools being asked
  different questions.
• test.sh is working-directory dependent, because mado's config is
  resolved from the cwd and nothing in the script sets one. Running the same
  binary over the same corpus gives 232 findings from the repository root and
  233 from /tmp. That is a reproducibility bug in the harness regardless of
  what the comparison eventually shows. Related to The development scripts report results without having run anything #400, which collects the
  other ways these scripts answer without having run what you think.

So the first thing to settle here is not "which tool is right" but "what should
the two configs be". Either test.sh passes both tools a matched pair written
for this purpose, or the difference between .mdlrc and mado.toml becomes
part of what is being compared, deliberately and in writing.