Compute each selector once, not seven times

mkdeps.sh asked the same questions over and over: "pages" is built
on four of the other selectors, the tagged list mkwb reads asks for
all seven again, and "collections" alone was computed nine times, a
find and a grep and a sort apiece.  It spawned 372 processes to
print 2450 lines.

Every answer is now computed once, up front, into a file named
after it, and read back from there.  The three hand-rolled caches
go with it: they were the same idea, done three times, and the
loop that fills them says the dependency order out loud.  Two
smaller things fell out on the way: collections ran dirname once
per collection where sed does it in one, and articles ran one find
per collection where find takes them all at once.

The order the loop fills them in is not decoration.  A pipeline's
stages run side by side, and only_masters and with_translations
are two stages of one selector that both read what po4a writes ---
so an answer computed where it is first needed can be read while
it is still half-written, which is a race that shows up as an
empty list and a deps.mk missing half the site.  Filled in order,
up front, nothing is ever read while it is being written.

372 processes to 109, 0.47s to 0.24s, byte-identical deps.mk over
a dozen runs.  The floor now is exec(2): the tools this reaches
for cost 1.1 to 2.5ms each to start, and a hundred of them is the
whole of the remaining runtime.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KxLpaNQThKfMG3ecCJptP9

0 files changed, 0 insertions, 0 deletions