gha memories: lint-markdown's fail: false does not cover its three companion checks

[d-morrison]

memories/gha-reusable-workflows.md records several consumer-side gotchas but
not this one, which cost a review round tonight.

lint-markdown.yml's fail input gates markdownlint only. The composite also
runs a list-item merge-splice check, a fenced-code-block length check, and a
GFM table-split check, gated by fail-on-item-splices,
fail-on-long-code-blocks, and fail-on-table-splits respectively --- each
defaulting to true.

So a caller written fail: false, to adopt the workflow warn-only while a
pre-existing backlog is worked down, still fails the build.

The reason it costs a round rather than a minute is that the evidence reads as
a contradiction: the caller says warn-only, the check is red, and the natural
hypotheses are a wrong pin or an ignored input. Both are wrong. A different
check inside the same job failed, and the job's conclusion comes from whichever
step failed --- which is the same shape fully-clean.md already describes for a
green guard step beside a red job.

lint-yaml.yml has the same split: fail gates yamllint, fail-on-long-scripts
gates its companion.

Measured on UCD-SERG/shigella#37. Its lint-markdown caller was added at
fail: false in UCD-SERG/shigella#33 and went red on six list-item splices.