Four more example decks, and move piper setup out of the README

[Aduneer]

Groundwork before v0.3.0, in three parts.

Example decks

Deck
http-status.csv	18 codes, tagged by class
elements.csv	19 symbols, weighted towards the ones from Latin names — the only ones worth a flashcard
nato-phonetic.csv	All 26
general-knowledge.csv	Grew from 5 cards to 12 real ones

What's 2 + 2 - 1? was filler next to every other deck here and is gone.

nato-phonetic.csv turns out to be the first shipped deck that exercises the single-key guard. Reviewed reversed its answers are single letters, so a card answered Q collides with quit and one answered A with audio. Checked against the real binary rather than assumed:

reversed, answer is "Q":   [Enter] submit   [a] play audio   [?] hint
reversed, answer is "A":   [Enter] submit   [?] hint   [q] end session
normal,   answer "Quebec": [Enter] submit   [a] play audio   [?] hint   [q] end session

Both withheld exactly when the card could accept them, and both answer correctly.

Two things I deliberately did not do

No pictures on the existing language decks. spanish.csv and japanese.csv run foreign → English, so a picture shows the meaning, which on those cards is the answer. That is the rule colores.csv exists to demonstrate, and following it here means not doing the obvious thing.

No shipped audio, for the reason --generate-audio exists: recordings are rendered locally, not committed. What was actually missing was the command, so the README now carries it — with voice names taken from voices that exist rather than guessed. ja_JA-hi_fi_captain-medium is not the ja_JP the naming convention suggests.

README

The piper install-and-troubleshoot material moves to docs/audio.md. It was the one section that was reference rather than tour: how to install a TTS engine, where its voices live, the extra package Japanese needs.

Being straight about the result: the README did not get shorter — 653 lines to 657. The setup appendix left; a deck table and the audio recipe arrived. That is a better balance rather than a smaller file. The bigger split (Syncing and Development out too, roughly halving it) is still available if it is ever wanted.

A test for content

The example decks were entirely uncovered. They are data, so a missing comma does not fail to build — it quietly drops a card, or ships one whose answer is empty and can therefore never be got right.

Every examples/*.csv is now loaded and checked: every non-blank line became a card, no question or answer is blank or normalises to nothing, and every picture a deck names is actually present.

Verified the check can fail rather than trusting that it passes:

$ printf 'git stash pop\n' >> examples/git.csv        # a card with no answer
FAIL (line 1901): deck.size()

$ sed -i '1s|images/red.png|images/missing.png|' examples/colores.csv
FAIL (line 1912): picture.good() was false

Testing

1184 unit checks (up from 680 — the deck validation checks every card) and 47 golden cases, green on gcc and clang, optimised and sanitised, clean in an env -i clean room, and verified from a fresh clone.