Some cookbooks style a recipe's first ingredient with its own CSS class
(eg. ingredient1) and the rest with another (ingredient). The labeller
only recognised the repeated 'run', so the first ingredient was dropped —
and recipes whose whole ingredient list used the first-ingredient class
were rejected for having no ingredients.

Fold in a once-per-recipe class that leads with a numeral (a quantity),
is ingredient-length rather than prose, and sits with the ingredient run.
The description-vs-method split still uses the run's position, so an
early-positioned first-ingredient class doesn't skew it.

Across the test library this recovered first ingredients in 464/502
recipes of one book and added 2 previously-empty recipes in another, with
no change to any book's instructions.

Cookbooks often give a recipe a native title and a translated/generic
name (Abricotines / Apricot Balls; Saag Paneer / Spinach with Indian
cheese). Store both on the interchange recipe (new Subtitle field) and
render them together via DisplayTitle as 'Title (Subtitle)', which every
output format now uses for its title.

ePub extraction fills Subtitle from the induced subtitle instead of
folding it into the description. A subtitle line that carries a number
('Makes 56') is a yield rather than a name, so it's recorded as the yield
— which also keeps it out of the description.

ePub recipes had no source, so the book they came from was lost. Record
it on the interchange recipe: the book title as the source name and its
ISBN as a urn:isbn URI. Mela maps the URI to its link and, having no
source field, attributes the book name in the notes.

Mela also renders newlines literally, so a description's paragraph breaks
('\n\n') showed as blank lines; collapse each run to a single newline
when importing to Mela.

Sel.Matches treated an empty Class as a wildcard, so a classless selector
like p matched every paragraph. When the labeller assigned a book's
classless prose paragraphs to the steps role, that selector then swallowed
the title, description and ingredients into the instructions. Treat an
empty Class as the classless group specifically.

When a book can't self-certify structurally, InduceWith re-labels every
candidate delimiter with the model labeller. A class's role, though, is
intrinsic to its text, not to the delimiter — yet Label re-embedded every
class's sample lines on each of the (up to 40) candidates, so the model
path ran for minutes and timed out.

Embed and classify each class once, cache the role by class selector, and
reuse it for later candidates; a mutex also serialises the non-reentrant
model session when books are induced concurrently. On a book that needs
the model, the induction drops from >2 min (timing out) to ~40s. The
embedding call is now injectable so the caching is unit-tested without
the model.

Two issues surfaced converting an image-rich ePub cookbook to Mela:

Photos were silently dropped. The jpegli package (imported for its
faster encoder) registers a WASM-backed JPEG decoder for the global
image registry, and it panics ("unreachable") on some otherwise-valid
JPEGs. Every recipe photo tripped it at the decode-config step, so all
were rejected and the size filter learned nothing. Decode JPEG/PNG with
the standard library directly (new utils.DecodeConfig/DecodeImage),
keeping jpegli for encoding only.

Descriptions could carry stray blank lines. An induced description joins
its text blocks with a blank line, so a block with its own trailing
newline stacked into three or more. Collapse any run of 2+ newlines (with
interspersed whitespace) to a single blank line and trim.

The grammar only accepted unicode fraction glyphs (½) and decimals, so a
plain ASCII quantity like "1/2" or "1 1/2" — common in older formats and
hand-typed recipes — was left unparsed and leaked into the ingredient
name. Add an AsciiFraction rule (and a mixed "integer space fraction"
form) to the Amount rule so every format now reads these as exact
rationals.

This supersedes the NormalizeAmount preprocessing helper: the MealMaster
and RecipeML importers now pass their raw quantity fields straight
through, and odd denominators (eg. 1/16) parse exactly rather than as a
rounded decimal.

RecipeML is an open early-2000s XML recipe format with a clean structure
(separate quantity / unit / item nodes); sizeable .xml archives remain
though the ecosystem is dead. A document may hold many <recipe> elements,
so Parse streams one event each and returns no collection details.

The parser maps head/title, description, categories, note, and a yield
rebuilt from <range>/<q> plus any trailing unit text. Ingredients become
structured groups — ungrouped <ing>s in a leading untitled group, each
<ing-div> its own titled group — reconstructing "qty unit item (prep)"
and folding <prep>/<modifier> into the note. Inline <step> markup is
flattened to text. Decoding honours a declared charset and never fetches
external DTDs (no XXE). .xml is claimed eagerly; non-RecipeML XML yields a
clear non-fatal error. Writing is not supported.

MealMaster's helper for turning ASCII fractions ("1/2", "1 1/2") into a
form the ingredient grammar accepts is needed by the RecipeML importer
too, which reconstructs an ingredient line from a separate quantity
field. Move it to internal/ingredients as NormalizeAmount.

MealMaster is a 1990s DOS recipe program whose plain-text export holds a
vast body of BBS-era recipe archives. A single .mmf usually concatenates
many recipes, so Parse streams one event per recipe found and returns no
collection details (it's a flat archive, not a named app collection).

The parser splits on the 'via Meal-Master' headers and hyphen/MMMMM
terminators, reads the Title/Categories/Yield fields, and decodes the
fixed-column ingredient block (amount cols 1-7, 2-letter unit cols 9-10,
text cols 12+) including continuation lines, hyphen-wrapped section
headers, and the side-by-side two-column layout. Ingredient lines are
reconstructed into grammar-friendly text — ASCII fractions become
unicode glyphs and unit codes expand to words — then parsed into
structured ingredients. Files are decoded from CP437 when not UTF-8.

Writing is not supported.