alg-viii: the 317 exercises split into one file each

[tamnd]

Bourbaki prints the exercises of a section as a run at its foot, one paragraph to an exercise. They are now one file to an exercise, under content/en/alg/VIII/exercises/sN for a § and exercises/aN for an appendix, because the solver takes its work one exercise at a time. Each section keeps its anchored ### Exercises heading and a link to the directory, so a reader following the book still lands where the run was printed.

21 sections and 3 of the 4 appendices have exercises, appendix 3 having none, which is 24 directories and 317 files. 76 are starred and 3 supplementary. An appendix numbers its exercises from one exactly as a § does, so appendix 1 and § 1 both open with an exercise 1 and the directory says which is which.

manifests/exercises.json comes with them, which is what the solver reads to pick its work. It is JSON and not YAML because it is the one manifest meant to be read by something other than a person. It records for every section the count, the first and last number, the gaps, and how many are starred or supplementary.

Pages 446, 465 and 471 were read again. A parenthesised number dropped below the line is an index, not a footnote reference, and extraction had been reading it as one, so X_{\sigma(1)}, v_{\sigma(1)} and \varepsilon_{\sigma(1)} had come out as X_{\sigma}[^1] and each then pointed at a real footnote of the § it stood in. All three are back.

Audited over the whole chapter before this was pushed. No gap in any section's numbering and the files run 01.md to NN.md contiguous from one. The counts in the manifest match the files on disk and the exercises: line of each section's front matter, 317 in total either way. Every one of the 317 agrees with the starred and supplementary markers on the page it opens on. Every section's count matches the highest number the pages print after its exercises heading. Every footnote definition has a mark in the same file and every file numbers its notes from one. No file is missing front matter and no file carries a copyright line.

One known extraction artifact is left, a stray [^1] inside an already mangled display in exercise 25 of § 21, from page 449. It no longer steals the note of § 21 and it belongs in the extraction quality report rather than in a rule nothing measured.

Written by assemble at github.com/tamnd/bourbaki-solver#24, which runs assemble -book alg-viii -check in CI here.

Part of milestone #5.