mcp-recipes 1.1.0 → 1.3.0

package/CHANGELOG.md

@@ -4,6 +4,276 @@ All notable changes to this project are documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and the versions follow
 [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## 1.3.0
+
+### Added
+
+- A question asked in a sentence is sent in shorter wordings as well as in the
+  words it was written in. These indexes answer the words they are handed: one
+  requires every word given to appear on the same page, so "comment faire un
+  ceviche de truite à l'origan" came back empty from a corpus holding four
+  ceviche recipes, and the other ranks on the words received, so the words
+  framing a question pulled up whatever shared them and pushed the dish out of
+  the list. A question now becomes an ordered ladder of at most three wordings,
+  tried in order and stopped as soon as enough rows have come back, and the rows
+  are the union deduplicated on the id naming its source. `per_source` lists
+  every wording, what each returned, and the ones held back with the reason.
+  Nothing is translated between the sources' languages.
+
+- `search_recipes` takes `fan_out`, on unless a caller turns it off, which sends
+  exactly the words typed and still names the wordings that were withheld.
+
+### Changed
+
+- A condition a question sets is handled as a condition rather than as a word to
+  find. "sans beurre", "gluten free" and "pour 6 personnes" stay in the wording
+  sent first, where a source whose index carries them can match, and are set
+  aside from the derived wordings, since no page is named for what it leaves out
+  and a number of eaters is not printed in a title. Every condition set aside is
+  named back in the notes with a reminder to read the ingredient list, because a
+  recipe found this way can hold exactly what was being avoided. A number of
+  servings is pointed at the `servings` argument.
+
+- A condition is read by the role its words play rather than by the turn of
+  phrase they are written in. Reading it off the phrasings "sans X", "without X"
+  and "X-free" left an allergen a person had declared to become a positive
+  search term: "je suis allergique aux noix" put "noix" into a derived wording,
+  which pulled a coconut cake to the second row of a lasagne search with nothing
+  said about nuts. An exclusion written "no butter" or "pas de beurre" raised no
+  warning at all, and a diet named in one word was not treated as a condition,
+  so "recette vegane au foie gras" returned foie gras at the first row unmarked.
+  Three roles are now read: a negation, whichever of the closed class of
+  negating words carries it; a sentence about the eater rather than the dish,
+  which is named back as an allergy because that is the condition whose failure
+  is a matter of health; and a diet named in one word. All three are set aside
+  from the derived wordings and named back in the notes. A question can still
+  phrase one in a way this reading misses, and the notes say so instead of
+  presenting the conditions read as the only ones present. The reading is made
+  once for the whole question, so the words set aside as a condition and the
+  words searched for as a dish can no longer disagree.
+
+- A quantity written across two measures is read as one quantity. "1 lb 4 oz
+  beef" doubled came back as "2 lb 4 oz beef", a tenth short of the answer and
+  marked as the exact product, and "1 kg 500 de farine" came back as "2 kg de
+  500 de farine". Two measures off the same ladder written side by side, with
+  the second holding less than one of the first, are one amount: they give
+  "2.5 lb beef" and "3 kg de farine". The French form that leaves the smaller
+  unit unwritten is read the same way.
+
+- A measure standing in front of a container is read as the size of one of them.
+  "12 oz can tomatoes" doubled came back as "1.5 lb can tomatoes", which asks
+  for a bigger tin where the recipe wants a second one. Lines of that shape come
+  back as published, with a note saying what the measure gives and that the line
+  states no count.
+
+- A further quantity on a line is seen where the measure stands behind the word
+  introducing it. "150 g de sucre (soit 3/4 de tasse)" doubled came back as
+  "300 g de sucre (soit 3/4 de tasse)" marked as exact, so the line claimed
+  300 g was three quarters of a cup. A quantity left as published beside a mass
+  or a volume now shows on the line's own `scaling`, which no longer reads
+  `scaled`. A container's stated capacity is not one of those: "2 boîtes de
+  400 g" is exactly twice "1 boîte de 400 g" and stays `scaled`.
+
+- A count read from a word that names no number is not reported as exact
+  arithmetic. "quelques feuilles de basilic" doubled gave six leaves marked
+  `scaled`, which put the weight of arithmetic behind a reading of "quelques".
+  The line comes back `rounded`, and the note says the figure is that reading
+  multiplied rather than a count the page gave.
+
+- `compare_recipes` says when no version carries a word of the dish in its
+  title. One of these indexes answers almost any spelling with its closest row,
+  so a dish nobody publishes came back as a doughnut recipe presented as a
+  source's version of it, closing on an invitation to read it as what that
+  source publishes. That invitation is now withheld in exactly that case, and
+  the answer says the rows are candidates to check.
+
+- A part of a recipe that comes back empty is reported as something this server
+  read nothing from rather than as something the page does not publish. Where
+  the page shows a sign of it anyway, a heading announcing that part or another
+  part of the recipe that was read, the answer states that this server failed to
+  read what the page carries. Where the page shows neither, the answer states
+  that a page publishing nothing and a layout this reader cannot follow are
+  indistinguishable from here. Both wordings close on the same warning: an empty
+  ingredient list is no evidence that an ingredient is absent from the dish.
+  This covers an empty method the same way. The note is only raised for a
+  section that was asked for.
+
+### Fixed
+
+- An allergy stated with the food before the marker is read. Scoping a condition
+  forwards alone reads the French order, "allergique aux noix", and misses the
+  English one, so "cookies peanut allergy" searched for the nut, returned peanut
+  butter and said nothing about an allergy; "pancakes, I am lactose intolerant"
+  did the same with lactose. A marker now takes its food from whichever side of
+  it the sentence put one: a joining word straight after it says the food
+  follows, and with no such word the food is the one the marker was written onto
+  the end of.
+
+- An exclusion and the dish no longer swap places. "a dessert free of dairy"
+  reported the dessert as the thing to leave out and searched for the dairy.
+  "free" is read the same way as any other marker, so "free of dairy" leaves out
+  the dairy and "gluten free" leaves out the gluten.
+
+- A condition ends at the food it names. A fixed span of words let a condition
+  written at the head of a question swallow the dish behind it, so "sans gluten
+  gateau au chocolat" reported a recipe free of "gluten gateau au". A food's
+  name now runs on only through a word that carries it into a second noun, as in
+  "farine de blé", and stops everywhere else.
+
+- A marker with a food on neither side of it is reported as a condition whose
+  food was not read, rather than passing unmentioned or taking a word at random.
+  Naming the wrong word hides a dish and searches for the food being avoided at
+  the same time, so the answer states that a condition was set and that this
+  server did not read what it covers.
+
+- The note that says a diet was not filtered on reaches the text block. Notes
+  are dropped from that block when an answer carries more than it has room for,
+  and which ones could go was decided by matching their wording, so the diet
+  note went with the rest and "recette vegane au foie gras" rendered pan-fried
+  foie gras at the first row with nothing said about the diet. A note now
+  carries whether the answer misleads without it, set where the note is written,
+  and the block gives up body text before it gives up such a note.
+
+- A comparison confronts two versions only where both titles carry the dish.
+  Sharing one word of a name is what makes a row worth offering and not what
+  makes it the dish, so "biscuits and gravy" put a tin of Christmas biscuits
+  beside the American dish and reported their yields, their times and their
+  ingredient counts as the difference between two traditions. A version whose
+  title leaves a word of the dish unanswered is now named, with the words it
+  does not carry and an invitation to read it as a candidate to check, and
+  nothing is set side by side until every version answers to the same name.
+
+- A bracket the line closes on is scaled with the amount it restates.
+  "150 g de sucre (soit 3/4 de tasse)" doubled came back with the bracket as
+  published, so the line claimed that 300 g is three quarters of a cup. A
+  closing bracket that reads entirely as measures now moves with the figure it
+  restates, and it is read through the word that introduces it and through the
+  partitive between the number and the measure, so "soit 3/4 de tasse" is the
+  same quantity as "0,75 tasse". A slash inside a bracket separates two
+  statements of one quantity, and one sitting between two digits belongs to the
+  fraction it writes.
+
+- A capacity in brackets counts containers rather than measuring one out.
+  "1 (14 oz) can" doubled came back as "2 (1.75 lb) can", which asks for a tin
+  no shop sells and converts a quantity into another system on the way. A
+  measure inside a bracket standing beside a container is what one of them
+  holds: it goes back exactly as published, the count is what the factor moves,
+  the container agrees with that count, and the line says which figure is
+  whose.
+
+- A bracket a French line closes on moves with the rest of the line.
+  "1 demi verre de lait (10 cl)" tripled instructed the reader to take the ten
+  centilitres as published while asking for three times the milk.
+
+- A measure written in abbreviations keeps what stands behind it.
+  "2 c.à.s d'huile" doubled came back as "4 cuillères à soupe", with the oil
+  gone from the recipe, and "2 c.à.s + 1 c.à.c d'huile" came back as
+  "4 cuillères à soupe de c.à.c d'huile". Normalising an abbreviation spells it
+  as several words where the line writes one, and that difference was consuming
+  words of the ingredient; the words the line wrote are now consumed until they
+  match the measure. Two measures joined by a sign or a word saying they add up
+  are read as one quantity, as two measures written side by side already were.
+
+- A mass keeps the precision the arithmetic gave it. "1234 g" doubled came back
+  as "2,47 kg" marked as the exact product, where the exact product is 2468 g:
+  a whole number of grams was being snapped to a multiple of five, and the
+  result was then read in kilos, where a hundredth of the unit is ten grams and
+  every gap looks small. A whole number is what a scale shows and is left where
+  the arithmetic put it, a quantity moves up to a bigger unit only where that
+  unit states it as precisely, and whether a value landed exactly is judged in
+  the finer of the two units.
+
+- A mark standing in front of the amount is not the amount. A line opening on a
+  bullet, a dash or a picture of the food was answered as a line carrying no
+  quantity, so 200 g of cheese stayed 200 g in a doubled recipe with nothing
+  said, and the mark also silenced the measure the language is read from. Such
+  a mark is set aside for reading and put back for writing.
+
+- A text block that runs out of room says what it left out. Sentences
+  qualifying an answer keep their room, which leaves the result lines as the
+  thing that gives way, and a list of 119 ingredients gave way to nothing at
+  all in silence, which reads as a recipe needing no ingredients. A list now
+  keeps its opening whatever the room allows, states how many of how many lines
+  are shown, and points at the structured output for the rest. A recipe's
+  ingredients and its method share what is left rather than the first of them
+  filling the block.
+
+- A length of time is no longer multiplied like an ingredient. Ingredient lists
+  carry a rest, a proof or a marinade among the quantities, and "2 h de repos"
+  tripled came back as "6 h de repos". A figure followed by a measure of time,
+  in either language, is returned as published with a note saying that a rest
+  takes as long for a large batch as for a small one.
+
+- A word that names no number no longer produces one. "quelques feuilles de
+  basilic" tripled came back as "9 feuilles de basilic", while "a few basil
+  leaves" and "plusieurs feuilles de basilic" were returned as published, so
+  three ways of writing the same vagueness got three answers. All three are read
+  the same way now: the line carries no count, so it is repeated as published
+  and flagged as carrying nothing to multiply.
+
+- The factor stated is the factor applied. It was shortened to three decimals
+  while the arithmetic was not, so `factor: 0.0001` rendered "Scaled by 0" and
+  put a zero in the payload beside quantities that were not zero. Below three
+  decimals the significant digits are stated instead.
+
+- One quantity gives one answer however the page spelled it. "1234 g de farine"
+  and "1 kg 234 de farine" doubled came back 32 g apart, one of them called
+  exact and the other rounded, because the kilo cannot state 2468 g in the two
+  decimals a kitchen reads. A measured amount now moves down to the unit that
+  states it exactly, the way it already moved up only to a unit that could hold
+  it, so both lines answer "2468 g de farine".
+
+- `compare_recipes` states differences about the parts it returned. It compared
+  the times of versions whose times the same answer said were not requested and
+  not returned, leaving a reader with a claim they had no way to check; a
+  difference is now stated only about a section the call asked for. Each
+  version's yield is also written the way the rest of the answer writes one, as
+  published and with the factor and the number of servings it was rescaled for,
+  rather than as a bare published yield standing above a rescaled list.
+
+- Four smaller repairs. A query carrying no letter and no digit is refused
+  rather than sent, since an index handed nothing to match on answers with the
+  page it shows by default and those rows came back as recipes for "!!!". A
+  failure raised while reading a source is restated in this server's own words,
+  with what the site said quoted and credited to it and with any advice it
+  carried dropped, rather than served as this server's own sentence pointing
+  somewhere this server cannot answer for. A figure glued to an ordinal ending
+  is a rank rather than an amount, so "1er choix de boeuf" doubled no longer
+  reads "2 er choix de boeuf". And a rewritten line keeps the case the page
+  used, "1 TOMATO" tripled giving "3 TOMATOES", while a French noun in -ou takes
+  the plural French gives it: three choux, not three chous.
+
+## [1.2.0] - 2026-08-07
+
+### Changed
+
+- A mass stated behind the thing counted is read as the size of one of them.
+  "1 dinde de 3 kg" names one bird and says what it weighs, and multiplying the
+  count answered a dinner for twelve with a bird and a half. Lines of that shape
+  come back as the page published them, with a note saying the figure gives a
+  size and that more people means a bigger bird. A container keeps counting:
+  "2 boîtes de 400 g de tomates" states what one boîte holds and goes on to name
+  what is in it, and an equivalence such as "450 g (1 livre) de spaghetti"
+  restates one quantity twice and is scaled on both sides.
+
+- A comparison tells apart the two moments a source can fail in. Searching and
+  then opening the version a search offered are two different readings, and a
+  source missing from a comparison was described as one whose search did not
+  answer whichever of the two had happened. Every row of `per_source` now
+  carries a `read` field saying what became of the version that source offered,
+  and the notes name the moment that failed.
+
+- A search Marmiton answers with nothing is counted as an answer. Marmiton
+  serves a 404 on its results page when no recipe matches, and reading that as a
+  failed source put "Marmiton did not answer" in front of a reader for a search
+  Marmiton had answered. Every other failure stays a failure, so a site that
+  cannot be reached is never read as a site that holds nothing.
+
+- A list of results carrying none of the words asked for says so. A source
+  ranking a title on the letters it opens with answers "chameau farci" with a
+  chapeau and a gâteau château. When no title returned carries a word of the
+  query, the answer names that, and the rows read as candidates to check.
+
 ## [1.1.0] - 2026-08-07
 
 ### Changed

package/README.md

@@ -1,5 +1,11 @@
 # mcp-recipes
 
+[![npm](https://img.shields.io/npm/v/mcp-recipes.svg)](https://www.npmjs.com/package/mcp-recipes)
+[![CI](https://github.com/smeet666/mcp-recipes/actions/workflows/ci.yml/badge.svg)](https://github.com/smeet666/mcp-recipes/actions/workflows/ci.yml)
+[![license](https://img.shields.io/npm/l/mcp-recipes.svg)](LICENSE)
+[![MCP Registry](https://img.shields.io/badge/MCP_Registry-listed-6E56CF)](https://registry.modelcontextprotocol.io/v0/servers?search=io.github.smeet666/mcp-recipes)
+[![Glama](https://glama.ai/mcp/servers/smeet666/mcp-recipes/badges/score.svg)](https://glama.ai/mcp/servers/smeet666/mcp-recipes)
+
 One question, several kitchens. An MCP server that asks every recipe source it
 reads at the same time, merges what they say, and scales quantities in French
 and English.
@@ -94,6 +100,17 @@ clou de girofle, and a count of those lands on a whole number.
   other is the page's own claim.
 - `4 tablespoons` at a tenth comes down to **1.2 teaspoons**. A measurement moves
   to a smaller unit _before_ it is rounded, so a small share survives.
+- `1 lb 4 oz beef` doubled gives **2.5 lb beef**, and `1 kg 500 de farine` gives
+  **3 kg de farine**. Two measures off one ladder written side by side are one
+  quantity, and the smaller one can be left unwritten.
+- `12 oz can tomatoes` comes back as published, with a note saying the measure
+  gives what one tin holds and that the line states no count of them.
+- `2 h de repos` comes back as published. A rest, a proof or a bake takes as long
+  for a large batch as for a small one, so a length of time is not a proportion.
+- `quelques feuilles de basilic` comes back as published, as do `a few basil
+leaves` and the same line with `plusieurs`. A word saying there are several
+  states no count, and any figure put behind it would be this server's reading
+  rather than the page's quantity.
 
 Every line comes back with a `scaling` field: `scaled` when the arithmetic
 landed on the exact product, `rounded` when something had to move for the line
@@ -145,12 +162,70 @@ one costs.
 | `query`            | string                        | A dish or an ingredient, in any language a source publishes in |
 | `limit_per_source` | integer, 1 to 25, default 5   | Rows to take from each source                                  |
 | `sources`          | array of source ids, optional | Left out, every source is asked                                |
+| `fan_out`          | boolean, default true         | Whether shorter wordings derived from the question go out too  |
 
 The query goes to each source's own search as free text. There is no filtering:
 a word naming a diet, a time or a calorie count matches only where that source's
 index happens to carry it, and a question long enough to read like a request for
 conditions gets a note saying so.
 
+**A question asked in a sentence is also sent in shorter wordings.** These
+indexes answer the words they are handed. One of them requires every word given
+to appear on the same page, so "comment faire un ceviche de truite à l'origan"
+comes back empty from a corpus holding four ceviche recipes; the other ranks on
+the words it was handed, so "comment faire" pulls up whatever shares those words
+and the dish falls out of the list. Both reach a reader looking like a statement
+about a corpus when they are statements about a wording.
+
+So a question is turned into an ordered ladder of at most three wordings: the
+words as asked, the words naming the dish with the framing words set aside, and
+the leading dish word on its own. They go out in that order, one at a time, and
+stop as soon as enough rows have come back, so a query that already names a dish
+costs one request. What comes back is unioned and deduplicated on the id naming
+its source. `per_source` lists every wording, what each returned, and the ones
+held back and why. Nothing is translated between the sources' languages: a
+guessed translation is a word nobody wrote, and a source that publishes in
+another language and holds nothing under the words written is an honest absence.
+
+**A condition is not a keyword.** "sans beurre", "gluten free" and "pour 6
+personnes" say what a recipe must avoid or serve. No page is named for what it
+leaves out, and the number of people at a table is not printed in a title, so
+searching for those words buries the dish. They stay in the wording sent first,
+where a source whose index carries them can still match, and they are set aside
+from the derived wordings. Dropping them quietly would be the worse move, since
+a recipe found this way can hold exactly what the person was avoiding, so every
+condition set aside is named back in the notes with a reminder to check the
+ingredient list. A number of servings is pointed at the `servings` argument,
+which is what actually rescales a recipe.
+
+A condition is read by the role its words play rather than by the turn of phrase
+they are written in, because a list of phrasings has a hole at the first one
+nobody thought of. Three roles put a food out of a dish: a negation, whatever
+word carries it, since negations are a closed class of function words and
+"without butter", "no butter" and "pas de beurre" are all built on one; a
+sentence about the eater rather than the dish, as in "je suis allergique aux
+noix", which is named back as an allergy because that is the condition whose
+failure is a matter of health; and a diet named in one word, which is a class of
+recipes defined by what it leaves out and so states an exclusion with no
+negation written anywhere. A question can still phrase one in a way this reading
+does not catch, and the answer says so rather than presenting the ones it caught
+as all there were.
+
+Which words a condition covers is read off the sentence. A joining word straight
+after the marker says its food follows, as in "allergique aux noix", "allergic
+to peanuts" or "free of dairy"; with no such word, the food is the one the
+marker was written onto the end of, as in "peanut allergy", "lactose intolerant"
+or "gluten free". Both orders are needed, since the two languages these sources
+publish in put the food on opposite sides. A food's name ends where nothing
+carries it further, so "sans gluten gateau au chocolat" leaves out gluten and
+looks for the cake. Where a marker has a food on neither side, the answer says a
+condition was stated and that its food was not read: naming the wrong word hides
+a dish and leaves the food being avoided in the search at the same time, and an
+allergen is not something to guess at.
+
+Setting `fan_out` to false sends exactly the words typed. The wordings that were
+not sent are still named, so a caller can see what turning it back on would buy.
+
 Rows are interleaved one source at a time. No score orders them against each
 other, because the sources share none: one carries reader ratings and another
 has no author and no rating by nature. The answer says how the order was built.
@@ -188,8 +263,17 @@ the lower end, and the answer gives the factor the upper end would have needed.
 
 Some sources keep recipes and reference pages under one namespace, so a search
 row can be a page _about_ a dish rather than a recipe _for_ it. Only reading the
-page tells them apart, and this tool says when a page carries no ingredient
-list.
+page tells them apart.
+
+A part of a recipe this answer holds nothing for is never reported as an absence
+on the page, because reading nothing is a fact about this server and only the
+page can say whether there was anything to read. Where the page shows a sign of
+the part all the same, a heading announcing it or another part of the recipe
+that was read, the answer states that this server failed to read what the page
+carries. Where the page shows neither, it states that the two cannot be told
+apart from here. Either way an empty ingredient list is never evidence that an
+ingredient is absent from the dish: someone avoiding butter has to open the page
+before ruling it out.
 
 ### `scale_ingredients`
 
@@ -466,6 +550,18 @@ d'œuf se compte entier, le blanc de poulet est une viande et se coupe en deux.
   vaut l'autre, c'est la page qui l'affirme.
 - `4 tablespoons` au dixième descend à **1.2 teaspoons**. Une mesure passe à une
   unité plus petite _avant_ d'être arrondie, pour qu'une petite part survive.
+- `1 lb 4 oz beef` doublé donne **2.5 lb beef**, et `1 kg 500 de farine` donne
+  **3 kg de farine**. Deux mesures d'une même échelle écrites côte à côte sont
+  une seule quantité, et la plus petite peut rester sous-entendue.
+- `12 oz can tomatoes` revient telle que publiée, avec la mention que la mesure
+  donne ce que tient une boîte et que la ligne n'énonce aucun nombre de boîtes.
+- `2 h de repos` revient telle que publiée. Un repos, une pousse ou une cuisson
+  durent autant pour une grande fournée que pour une petite : une durée n'est pas
+  une proportion.
+- `quelques feuilles de basilic` revient telle que publiée, comme `a few basil
+leaves` et comme la même ligne avec `plusieurs`. Un mot qui dit qu'il y en a
+  plusieurs n'énonce aucun compte, et tout chiffre placé derrière lui serait une
+  lecture de ce serveur plutôt qu'une quantité de la page.
 
 Chaque ligne revient avec un champ `scaling` : `scaled` quand le calcul est tombé
 juste, `rounded` quand quelque chose a dû bouger pour que la ligne reste
@@ -504,17 +600,78 @@ npx -y mcp-recipes
 Interroge toutes les sources en parallèle : les demander toutes coûte à peu près
 ce que coûte en demander une.
 
-| Argument           | Type                                        | Sens                                                       |
-| ------------------ | ------------------------------------------- | ---------------------------------------------------------- |
-| `query`            | string                                      | Un plat ou un ingrédient, dans une des langues des sources |
-| `limit_per_source` | entier, 1 à 25, défaut 5                    | Lignes prises à chaque source                              |
-| `sources`          | tableau d'identifiants de source, optionnel | Absent, toutes sont interrogées                            |
+| Argument           | Type                                        | Sens                                                                |
+| ------------------ | ------------------------------------------- | ------------------------------------------------------------------- |
+| `query`            | string                                      | Un plat ou un ingrédient, dans une des langues des sources          |
+| `limit_per_source` | entier, 1 à 25, défaut 5                    | Lignes prises à chaque source                                       |
+| `sources`          | tableau d'identifiants de source, optionnel | Absent, toutes sont interrogées                                     |
+| `fan_out`          | booléen, défaut true                        | Envoyer aussi des formulations plus courtes dérivées de la question |
 
 La requête part vers la recherche de chaque source en texte libre. Il n'y a
 aucun filtrage : un mot désignant un régime, un temps ou un nombre de calories
 ne compte que si l'index de la source le porte, et une question assez longue
 pour ressembler à une demande de conditions reçoit une note qui le dit.
 
+**Une question posée en phrase part aussi sous des formulations plus courtes.**
+Ces index répondent aux mots qu'on leur donne. L'un exige que chaque mot fourni
+figure sur la même page : « comment faire un ceviche de truite à l'origan » en
+revient vide alors que le corpus tient quatre recettes de ceviche. L'autre
+classe sur les mots reçus : « comment faire » remonte tout ce qui les partage et
+le plat sort de la liste. Dans les deux cas, le lecteur reçoit ce qui ressemble
+à un constat sur un corpus alors que c'est un constat sur une formulation.
+
+Une question devient donc une échelle ordonnée d'au plus trois formulations :
+les mots tels quels, les mots qui nomment le plat une fois écartés ceux qui
+encadrent la question, puis le premier mot du plat seul. Elles partent dans cet
+ordre, une à la fois, et s'arrêtent dès qu'il y a assez de lignes : une requête
+qui nomme déjà un plat coûte une requête. Ce qui revient est réuni et dédupliqué
+sur l'identifiant qui nomme sa source. `per_source` liste chaque formulation, ce
+qu'elle a rendu, et celles qui ont été retenues avec la raison. Rien n'est
+traduit d'une langue de source vers l'autre : une traduction devinée est un mot
+que personne n'a écrit, et une source qui publie dans une autre langue et ne
+tient rien sous les mots écrits est une absence honnête.
+
+**Une contrainte n'est pas un mot-clé.** « sans beurre », « gluten free » et
+« pour 6 personnes » disent ce que la recette doit éviter ou servir. Aucune page
+n'est nommée d'après ce qu'elle exclut, et le nombre de convives ne figure pas
+dans un titre : chercher ces mots enterre le plat. Ils restent dans la
+formulation envoyée en premier, où une source dont l'index les porte peut encore
+les faire correspondre, et ils sont écartés des formulations dérivées. Les
+laisser tomber en silence serait pire, puisque la recette trouvée peut contenir
+exactement ce que la personne évitait : chaque contrainte écartée est donc
+rappelée dans les notes, avec l'invitation à vérifier la liste d'ingrédients. Un
+nombre de parts est renvoyé vers l'argument `servings`, qui est ce qui remet
+réellement une recette à l'échelle.
+
+Une contrainte se lit au rôle que jouent ses mots, pas à la tournure dans
+laquelle ils sont écrits : une liste de tournures se troue au premier cas non
+prévu. Trois rôles sortent un aliment du plat. Une négation, quel que soit le
+mot qui la porte, puisque les négations forment une classe fermée de mots
+grammaticaux et que « sans beurre », « no butter » et « pas de beurre » reposent
+toutes sur l'une d'elles. Une phrase sur la personne plutôt que sur le plat,
+comme « je suis allergique aux noix », rappelée comme une allergie parce que
+c'est la contrainte dont l'échec relève de la santé. Et un régime nommé en un
+mot, qui est une classe de recettes définie par ce qu'elle exclut et énonce donc
+une exclusion sans qu'aucune négation soit écrite. Une question peut encore
+formuler une contrainte que cette lecture ne saisit pas, et la réponse le dit
+plutôt que de présenter les contraintes lues comme les seules présentes.
+
+La portée d'une contrainte se lit dans la phrase. Un mot de liaison juste après
+le marqueur dit que l'aliment le suit : « allergique aux noix », « allergic to
+peanuts », « free of dairy ». Sans ce mot, l'aliment est celui sur lequel le
+marqueur a été accolé : « peanut allergy », « lactose intolerant », « gluten
+free ». Les deux ordres sont nécessaires, puisque les deux langues que publient
+ces sources placent l'aliment de part et d'autre. Le nom d'un aliment s'arrête
+là où plus rien ne le prolonge, si bien que « sans gluten gateau au chocolat »
+écarte le gluten et cherche le gâteau. Quand un marqueur n'a d'aliment d'aucun
+côté, la réponse dit qu'une contrainte a été énoncée et que son aliment n'a pas
+été lu : nommer le mauvais mot cache un plat et laisse en même temps dans la
+recherche l'aliment que l'on évite, et un allergène ne se devine pas.
+
+Mettre `fan_out` à false envoie exactement les mots saisis. Les formulations non
+envoyées sont quand même nommées, pour qu'un appelant voie ce que le réactiver
+lui apporterait.
+
 Les lignes alternent d'une source à l'autre. Aucune note ne les classe les unes
 contre les autres, faute d'échelle commune : l'une porte des avis de lecteurs,
 l'autre n'a par nature ni auteur ni note. La réponse dit comment l'ordre a été
@@ -552,6 +709,16 @@ disant pourquoi. Une page qui annonce une fourchette, « 4 à 6 personnes », es
 recalculée depuis la borne basse, et la réponse donne le facteur qu'aurait
 demandé la borne haute.
 
+Une partie de recette dont cette réponse ne tient rien n'est jamais rapportée
+comme une absence sur la page : n'avoir rien lu est un fait sur ce serveur, et
+seule la page peut dire s'il y avait quelque chose à lire. Quand la page en
+montre malgré tout un signe, un titre de section qui l'annonce ou une autre
+partie de la recette qui a été lue, la réponse dit que ce serveur n'a pas su
+lire ce que la page porte. Quand la page n'en montre aucun, elle dit que les
+deux cas sont indiscernables d'ici. Dans les deux cas, une liste d'ingrédients
+vide ne prouve jamais qu'un ingrédient est absent du plat : qui évite le beurre
+doit ouvrir la page avant de l'écarter.
+
 ### `scale_ingredients`
 
 Hors ligne. Rien n'est téléchargé. La liste peut venir de n'importe quelle