@google-cloud/kcmd 0.0.0-stage → 0.1.0

package/build/ts/kcmd/semantic/osi_converter.js

@@ -0,0 +1,911 @@
+"use strict";
+// OSI (open, AI-first semantics) <-> Semantic Model IR converter.
+//
+// SCAFFOLD (naming for the end state): this file currently holds only the
+// WRITE direction -- IR -> OSI YAML (`serializeModel`). The READ direction
+// (OSI YAML -> IR) still lives in `loader.ts` and moves here once PR4 (the
+// KC push line) merges, at which point this becomes the full two-way
+// converter. Until then, "where is the OSI parser?" -> `loader.ts`.
+//
+// TODO(#278): fold the OSI READ direction in and drop the scaffold. Once
+// #278 merges, move `loadModels` / `fromDocument` / `loadSemanticModels`
+// (and their Load* types) out of `loader.ts` into this file, delete
+// `loader.ts`, and repoint its importers (`src/tool/commands.ts` and the
+// load / validate / bigquery tests). Then this is the sole OSI<->IR codec
+// and the SCAFFOLD note above comes out.
+//
+// Serializing is the inverse of `loader.ts`: `loader` reads authored YAML
+// into the IR; this module writes the IR out as a YAML document the loader
+// can read back. It is the local-workspace sink for `pull` (Knowledge
+// Catalog -> IR -> YAML), the counterpart to how `push` compiles YAML -> IR
+// -> a destination.
+//
+// Fidelity is at the IR level, not byte-for-byte with a hand-authored file. The
+// loader keeps every entry in `dialects` and records `stringForm`, and also
+// picks a target/canonical `expression` + an `importedExpression` via
+// `pickDialect`:
+//   * this serializer writes `dialects` back out when present (or short-form
+//     bare strings in GOOGLE_VERSION when `stringForm` is true), and falls back
+//     to `expression` and `importedExpression` when `dialects` is absent.
+//   * comments and key ordering are not preserved.
+// What the loader DOES keep on the IR round-trips here: names, descriptions,
+// `ai_context` (instructions / synonyms / examples / custom properties), keys
+// and unique keys, data sources, field datatypes (omitting `Opaque`) / labels /
+// dimension flags, expressions, relationship join columns, and
+// `custom_extensions`. Output format follows `model.version` via
+// `FORMAT_VERSION_PLACEMENT`: GOOGLE_VERSION ('0.2.0.dev0/google' or unset)
+// writes native extension keys and one `<model>.profile.<name>.yaml` sidecar
+// file per profile, whereas OSSIE_VERSION ('0.2.0.dev0') writes strict Ossie
+// keys, encodes `profiles`, `deployments`, `extends`, relationship
+// `description`, and metric `authoredEntity` in `GOOGLE` `custom_extensions`
+// blocks, and writes no profile files.
+//
+// An `association` (junction-table) relationship has no open-format syntax (the
+// loader cannot produce one), so only its direct foreign-key view (from/to +
+// columns) is serialized; the junction detail is dropped with a note.
+var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    var desc = Object.getOwnPropertyDescriptor(m, k);
+    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
+      desc = { enumerable: true, get: function() { return m[k]; } };
+    }
+    Object.defineProperty(o, k2, desc);
+}) : (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    o[k2] = m[k];
+}));
+var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
+    Object.defineProperty(o, "default", { enumerable: true, value: v });
+}) : function(o, v) {
+    o["default"] = v;
+});
+var __importStar = (this && this.__importStar) || (function () {
+    var ownKeys = function(o) {
+        ownKeys = Object.getOwnPropertyNames || function (o) {
+            var ar = [];
+            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
+            return ar;
+        };
+        return ownKeys(o);
+    };
+    return function (mod) {
+        if (mod && mod.__esModule) return mod;
+        var result = {};
+        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
+        __setModuleDefault(result, mod);
+        return result;
+    };
+})();
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.serializeModelFiles = serializeModelFiles;
+exports.serializeModel = serializeModel;
+exports.modelDocument = modelDocument;
+exports.compact = compact;
+exports.nonEmpty = nonEmpty;
+const yaml = __importStar(require("yaml"));
+const deployment_target_1 = require("./deployment_target");
+const ir_1 = require("./ir");
+const loader_1 = require("./loader");
+const names_1 = require("./names");
+const resolve_inheritance_1 = require("./resolve_inheritance");
+const resolve_profiles_1 = require("./resolve_profiles");
+// The vendor tag for Google-specific extension blocks on the IR (kept in sync
+// with loader.GOOGLE_VENDOR). In the Google flavor, a model-level deployment
+// target list rides in one as a folded shim alongside `model.deployments`.
+const GOOGLE_VENDOR = 'GOOGLE';
+// The dialect label for the IR's target/canonical `expression` when `dialects`
+// is absent and `stringForm` is not set. BIGQUERY is the loader's default
+// target dialect, so labeling the canonical form BIGQUERY makes the loader
+// re-pick it exactly on reload.
+const CANONICAL_DIALECT = 'BIGQUERY';
+// Only dialect labels the loader accepts are ever written out.
+const ALLOWED_DIALECT_SET = new Set(ir_1.ALLOWED_DIALECTS);
+/**
+ * Serializes a semantic model and its sidecar profile files. The returned
+ * `files` list always starts with `<model.name>.yaml`. Under `GOOGLE_VERSION`,
+ * each entry in `model.profiles` follows as
+ * `<model.name>.profile.<profile.name>.yaml`, except that a reserved
+ * (`default`), invalid, or case-insensitively repeated name is skipped with a
+ * warning. Under `OSSIE_VERSION`, valid profiles are written into the model's
+ * `GOOGLE` `custom_extensions` block and no profile files are emitted.
+ */
+function serializeModelFiles(model, opts = {}) {
+    const warnings = [];
+    const version = formatVersionOf(model);
+    const document = modelDocument(model, warnings, opts.logical);
+    const files = [
+        { fileName: `${model.name}.yaml`, content: render(document, opts) },
+    ];
+    if (loader_1.FORMAT_VERSION_PLACEMENT[version].model.profiles === 'file') {
+        for (const { name, doc } of profileDocuments(model, warnings)) {
+            files.push({
+                fileName: `${model.name}.profile.${name}.yaml`,
+                content: render(doc, opts),
+            });
+        }
+    }
+    return { files, warnings: [...new Set(warnings)] };
+}
+/**
+ * Returns the format version to serialize `model` as (`model.version`,
+ * defaulting to `GOOGLE_VERSION` when unset).
+ */
+function formatVersionOf(model) {
+    return model.version === loader_1.OSSIE_VERSION ? loader_1.OSSIE_VERSION : loader_1.GOOGLE_VERSION;
+}
+/**
+ * Validates `model.profiles` and returns the serialized `profileDocument`
+ * mapping for each valid profile.
+ *
+ * Reserved (`default`), invalid, and case-insensitively repeated names are
+ * skipped with a warning, keeping the first spelling, so the output satisfies
+ * the same name rules the profile-file layout and `readBlockProfiles` enforce
+ * on load.
+ */
+function profileDocuments(model, warnings) {
+    const seen = new Map();
+    const out = [];
+    for (const profile of model.profiles ?? []) {
+        const lower = profile.name.toLowerCase();
+        if (lower === resolve_profiles_1.DEFAULT_PROFILE) {
+            warnings.push(`model '${model.name}': profile '${profile.name}' uses the name ` +
+                `reserved for the model's inline bindings and is not serialized`);
+            continue;
+        }
+        if ((0, names_1.nameProblem)('profile', profile.name)) {
+            warnings.push(`model '${model.name}': profile '${profile.name}' is not a valid ` +
+                `profile name and is not serialized`);
+            continue;
+        }
+        const first = seen.get(lower);
+        if (first !== undefined) {
+            // Only a case-only repeat needs the ignoring-case rule explained.
+            const repeat = first === profile.name ?
+                `profile '${first}' is listed more than once` :
+                `profiles '${first}' and '${profile.name}' name the same ` +
+                    `profile (profile names are unique ignoring case)`;
+            warnings.push(`model '${model.name}': ${repeat}; only '${first}' is serialized`);
+            continue;
+        }
+        seen.set(lower, profile.name);
+        out.push({
+            name: profile.name,
+            doc: profileDocument(profile, warnings),
+        });
+    }
+    return out;
+}
+/**
+ * Serializes a single semantic model to a YAML document string in the open
+ * AI-first semantics format. The document contains exactly this one model;
+ * `pull` writes one file per model
+ * (catalog/EntryGroups/<entryGroup>/<model>.yaml). It is the first file
+ * `serializeModelFiles` returns, with the warnings for every file it builds.
+ *
+ * Warnings flag IR content that has no loadable representation (an association
+ * relationship's junction detail, an imported expression in an unknown dialect,
+ * dropped `GOOGLE` blocks under GOOGLE_VERSION, Google-only constructs and
+ * missing join columns under OSSIE_VERSION, or loader-invalid gaps), so the
+ * caller can surface the lossy edge.
+ *
+ * `logical` marks the model as a purely logical one (no physical binding).
+ * Under GOOGLE_VERSION it suppresses the missing-source and missing-expression
+ * warnings, which flag a lossy pull of a bound model. OSSIE_VERSION requires
+ * both, so they are still given. An OWL import sets it, and so does a
+ * second-generation pull, which loses no binding.
+ *
+ * `compactFlow` renders leaf collections (scalar sequences, all-scalar maps) in
+ * flow style so the output matches the compact convention the semantic-model
+ * guides author by hand and is byte-for-byte reproducible there. An OWL import
+ * sets it; `pull` does not.
+ */
+function serializeModel(model, opts = {}) {
+    const { files, warnings } = serializeModelFiles(model, opts);
+    return { yaml: files[0].content, warnings };
+}
+// Renders one document as YAML. `aliasDuplicateObjects: false` writes an IR
+// array that two keys share (an entity's key reused as a relationship's
+// `to_columns`, say) out in full at each key, rather than as an `&a1` anchor
+// and `*a1` alias, so editing one binding never silently edits another.
+// `compat: 'yaml-1.1'` quotes strings that a YAML 1.1 parser (PyYAML,
+// ruamel's 1.1 mode) would read as booleans or dates (`'yes'`, `'on'`,
+// `'2001-12-14'`), so arbitrary values in `custom_extensions` or `ai_context`
+// round-trip across tools.
+function render(doc, opts) {
+    return opts.compactFlow ? renderCompact(doc) : yaml.stringify(doc, {
+        aliasDuplicateObjects: false,
+        compat: 'yaml-1.1',
+        lineWidth: opts.lineWidth ?? 80,
+    });
+}
+// Renders "leaf" collections in flow style so the output matches the compact
+// convention the semantic-model guides author by hand: a sequence whose items
+// are all scalars (`primary_key: [id]`, `from_columns: [fk]`) and a map whose
+// values are all scalars (an inline field `{name, datatype}` or a column-less
+// relationship `{name, from, to}`). Nested structures -- a dataset, an
+// expression's `dialects`, an `ai_context` -- always stay block.
+//
+// A leaf collection is inlined ONLY while it stays short: a map or sequence
+// whose flow form would exceed FLOW_WIDTH_BUDGET (a field carrying a long
+// `description`, say) is left block, so `--compact` never emits an unreadable
+// multi-hundred-character line. `lineWidth: 0` disables the wrapper so an
+// inlined collection is never broken mid-braces -- the width gate, not the
+// wrapper, is what bounds line length. Flow and block parse identically, so
+// this only affects layout.
+const FLOW_WIDTH_BUDGET = 80;
+// The rendered width of a scalar node (its stringified value length); 0 for a
+// non-scalar, which never contributes because a collection holding one is not
+// inlined in the first place.
+function scalarWidth(node) {
+    return yaml.isScalar(node) ?
+        String(node.value ?? '').length :
+        0;
+}
+function renderCompact(document) {
+    const doc = new yaml.Document(document, { aliasDuplicateObjects: false, compat: 'yaml-1.1' });
+    yaml.visit(doc, {
+        Seq(_, node) {
+            if (!node.items.every(item => yaml.isScalar(item)))
+                return;
+            // `[` + each `item, ` + `]`, approximated for the width gate.
+            const width = node.items.reduce((n, item) => n + scalarWidth(item) + 2, 2);
+            if (width <= FLOW_WIDTH_BUDGET)
+                node.flow = true;
+        },
+        Map(_, node) {
+            // The document's own mapping stays block, as a hand-written file has it.
+            if (node === doc.contents)
+                return;
+            const pairs = node.items;
+            if (!pairs.every(pair => yaml.isScalar(pair.value)))
+                return;
+            // `{` + each `key: value, ` + `}`, approximated for the width gate.
+            const width = pairs.reduce((n, pair) => n + scalarWidth(pair.key) + scalarWidth(pair.value) + 4, 2);
+            if (width <= FLOW_WIDTH_BUDGET)
+                node.flow = true;
+        },
+    });
+    return doc.toString({ flowCollectionPadding: false, lineWidth: 0 });
+}
+// Builds the plain document object (version + one model) that yaml.stringify
+// renders. Kept separate so tests can assert the structure without parsing
+// YAML.
+function modelDocument(model, warnings = [], logical = false) {
+    const version = formatVersionOf(model);
+    return {
+        version,
+        semantic_model: [modelDoc(model, warnings, logical, version)],
+    };
+}
+// Builds a sidecar profile document matching the schema `loadProfileFile` in
+// `resolve_profiles.ts` reads: a top-level mapping with no `version` or
+// `semantic_model` wrapper. Optional lists (`primary_key`, `unique_keys`,
+// `fields`, `fields_exclude`, `metrics_exclude`, `actions`) are written
+// whenever defined, even when empty, so a ProfileSpec that loadProfileFile read
+// comes back the same from this file.
+function profileDocument(profile, warnings) {
+    return compact({
+        name: profile.name,
+        entities: nonEmpty(profile.entities.map(pe => compact({
+            name: pe.name,
+            source: pe.source,
+            primary_key: pe.primaryKey,
+            unique_keys: pe.uniqueKeys,
+            fields: pe.fields?.map(pf => {
+                const where = `profile '${profile.name}': entity '${pe.name}': field '${pf.name}'`;
+                const expression = expressionDoc(pf, /* isGoogleVersion= */ true, where, warnings);
+                // loadProfileFile rejects an entry with nothing to bind, and
+                // (like a model load) any blank SQL text.
+                const texts = typeof expression === 'string' ?
+                    [expression] :
+                    (expression?.dialects ?? []).map(d => d.expression);
+                if (expression === undefined) {
+                    warnings.push(`${where}: no expression; the profile file will not load ` +
+                        `until it is set or the field is listed in ` +
+                        `'fields_exclude'.`);
+                }
+                else if (texts.some(t => t.trim() === '')) {
+                    warnings.push(`${where}: expression is blank; the profile file will not ` +
+                        `load until its SQL is written.`);
+                }
+                return compact({ name: pf.name, expression });
+            }),
+            fields_exclude: pe.fieldsExclude,
+        }))),
+        relationships: nonEmpty(profile.relationships.map(pr => compact({
+            name: pr.name,
+            from_columns: nonEmpty(pr.fromColumns),
+            to_columns: nonEmpty(pr.toColumns),
+        }))),
+        metrics_exclude: profile.metricsExclude,
+        actions: profile.actions?.map(pa => compact({
+            name: pa.name,
+            executor: pa.executor === null ?
+                null :
+                (pa.executor ? executorDoc(pa.executor) : undefined),
+        })),
+    });
+}
+function modelDoc(model, warnings, logical, version) {
+    const isGoogleVersion = version === loader_1.GOOGLE_VERSION;
+    const placement = loader_1.FORMAT_VERSION_PLACEMENT[version];
+    // `datasets` is required (min 1) by the loader. A reconstructed model with no
+    // entities (e.g. every entity fetch failed during a pull) would serialize to
+    // a document the loader rejects; emit the (empty) array but flag it so the
+    // lossy edge is visible rather than surfacing later as an opaque load error.
+    const datasets = (model.entities ??
+        []).map(e => datasetDoc(e, warnings, logical, version));
+    if (!datasets.length) {
+        warnings.push(`model '${model.name}': no datasets (entities); the document requires ` +
+            `at least one and will not load until an entity is present.`);
+    }
+    const tableKey = isGoogleVersion ? 'entities' : 'datasets';
+    // Under GOOGLE_VERSION, deployments come from `model.deployments`, or from a
+    // GOOGLE `deploymentTargets` block when that is empty. Under OSSIE_VERSION,
+    // named deployments and profiles go in a GOOGLE block on the model.
+    const namedDeployments = (model.deployments ?? []).map(d => compact({
+        name: d.name,
+        target: d.target,
+        profile: d.profile
+    }));
+    const deployments = placement.model.deployments === 'native' ?
+        nonEmpty(namedDeployments.length ? namedDeployments :
+            deploymentsFromGoogleExtensions(model)) :
+        undefined;
+    const blockProfiles = placement.model.profiles === 'block' ?
+        nonEmpty(profileDocuments(model, warnings).map(p => p.doc)) :
+        undefined;
+    const blockDeployments = placement.model.deployments === 'block' ?
+        nonEmpty(namedDeployments) :
+        undefined;
+    const blockPayload = compact({
+        profiles: blockProfiles,
+        deployments: blockDeployments,
+    });
+    if (!isGoogleVersion)
+        warnGoogleOnlyContent(model, warnings);
+    const emittedFields = placement.model.actions === 'native' && model.actions?.length ?
+        emitted(model) :
+        null;
+    return compact({
+        name: model.name,
+        description: model.description,
+        ai_context: aiContextDoc(model.aiContext, isGoogleVersion),
+        deployments,
+        [tableKey]: datasets,
+        relationships: nonEmpty((model.relationships ??
+            []).map(r => relationshipDoc(r, warnings, version))),
+        metrics: nonEmpty((model.metrics ?? []).map(m => metricDoc(m, warnings, version))),
+        // Hoisted out of the map: `emitted` resolves inheritance, which deep-clones
+        // the model, and the answer does not vary by action.
+        actions: placement.model.actions === 'native' ?
+            nonEmpty((model.actions ??
+                []).map(a => actionDoc(a, warnings, emittedFields))) :
+            undefined,
+        constraints: placement.model.constraints === 'native' ?
+            nonEmpty((model.constraints ?? []).map(c => constraintDoc(c))) :
+            undefined,
+        custom_extensions: customExtensionsDoc(model.customExtensions, `model '${model.name}'`, isGoogleVersion, 
+        /* isModel= */ true, warnings, /* isField= */ false, blockPayload),
+    });
+}
+// Warns about each IR construct OSSIE_VERSION has no representation for,
+// natively or in a GOOGLE block, so a '0.2.0.dev0' write flags dropped
+// content just as a GOOGLE_VERSION write flags a dropped GOOGLE block.
+function warnGoogleOnlyContent(model, warnings) {
+    const placement = loader_1.FORMAT_VERSION_PLACEMENT[loader_1.OSSIE_VERSION];
+    const drop = (where, what) => warnings.push(`${where}: ${what} has no representation under '${loader_1.OSSIE_VERSION}' and ` +
+        `is not serialized`);
+    const where = `model '${model.name}'`;
+    if (placement.model.actions === 'none' && model.actions?.length) {
+        drop(where, `'actions'`);
+    }
+    if (placement.model.constraints === 'none' && model.constraints?.length) {
+        drop(where, `'constraints'`);
+    }
+    for (const e of model.entities ?? []) {
+        if (placement.dataset.abstract === 'none' && e.abstract) {
+            drop(`entity '${e.name}'`, `'abstract'`);
+        }
+    }
+}
+function datasetDoc(entity, warnings, logical, version) {
+    const isGoogleVersion = version === loader_1.GOOGLE_VERSION;
+    const placement = loader_1.FORMAT_VERSION_PLACEMENT[version];
+    // A concrete (non-abstract) entity with no source cannot be reloaded -- the
+    // loader requires `source` unless the dataset is abstract -- so surface the
+    // gap at write time instead of emitting a document that fails to load. Under
+    // OSSIE_VERSION `abstract` is omitted, so a dataset without `dataSource`
+    // cannot reload even if `entity.abstract` was set. Suppressed for a logical
+    // model under GOOGLE_VERSION, where a missing source is intended.
+    // OSSIE_VERSION requires a source on every dataset.
+    if ((!logical || !isGoogleVersion) && !(isGoogleVersion && entity.abstract) &&
+        !entity.dataSource) {
+        warnings.push(`entity '${entity.name}' has no source${isGoogleVersion ? ' and is not abstract' : ''}; the ` +
+            `serialized document will not reload until a source is set`);
+    }
+    const parents = nonEmpty(entity.extends);
+    const nativeExtends = placement.dataset.extends === 'native' ? parents : undefined;
+    const nativeAbstract = placement.dataset.abstract === 'native' ?
+        (entity.abstract || undefined) :
+        undefined;
+    const blockPayload = placement.dataset.extends === 'block' ?
+        compact({ extends: parents }) :
+        undefined;
+    return compact({
+        name: entity.name,
+        // An abstract entity has no physical table, so its source is empty; omit
+        // the key rather than emit `source: ""` (which the loader rejects). The
+        // source as written survives serialization; a legacy bare name is written
+        // as the name the load qualified it to.
+        source: (0, loader_1.writtenSource)(entity) || undefined,
+        // Supertype entities (entity-level inheritance); native under
+        // GOOGLE_VERSION, or in the dataset's GOOGLE block under OSSIE_VERSION.
+        extends: nativeExtends,
+        // Conceptual (table-less) marker; GOOGLE_VERSION only.
+        abstract: nativeAbstract,
+        primary_key: nonEmpty(entity.keys),
+        unique_keys: nonEmpty(entity.uniqueKeys),
+        description: entity.description,
+        ai_context: aiContextDoc(entity.aiContext, isGoogleVersion),
+        fields: nonEmpty((entity.fields ?? [])
+            .map(f => fieldDoc(f, warnings, logical || (isGoogleVersion && !!entity.abstract), isGoogleVersion))),
+        custom_extensions: customExtensionsDoc(entity.customExtensions, `entity '${entity.name}'`, isGoogleVersion, 
+        /* isModel= */ false, warnings, /* isField= */ false, blockPayload),
+    });
+}
+// Filters out the IR's `'Opaque'` placeholder so untyped fields and metrics
+// omit the `datatype` key in YAML. `'Opaque'` marks an unknown or unmapped type
+// (e.g. from `kcmd pull` or OWL import); action parameter helpers also use this
+// to tell whether a projected field will emit a concrete datatype on reload.
+function datatypeDoc(type) {
+    return type === undefined || type === 'Opaque' ? undefined :
+        type;
+}
+function fieldDoc(field, warnings, logical, isGoogleVersion) {
+    const expression = expressionDoc(field, isGoogleVersion, `field '${field.name}'`, warnings);
+    // Under GOOGLE_VERSION a field with no expression reloads unbound. A logical
+    // field, or one on an abstract entity, has none on purpose and is not warned.
+    // OSSIE_VERSION requires an expression on every field.
+    if ((!logical || !isGoogleVersion) && !expression) {
+        warnings.push(isGoogleVersion ?
+            `field '${field.name}': no expression; it reloads as an unbound ` +
+                `field.` :
+            `field '${field.name}': no expression; a '${loader_1.OSSIE_VERSION}' ` +
+                `document requires one per field and will not load until it ` +
+                `is set.`);
+    }
+    return compact({
+        name: field.name,
+        expression,
+        datatype: datatypeDoc(field.type),
+        label: field.label,
+        dimension: dimensionDoc(field),
+        description: field.description,
+        ai_context: aiContextDoc(field.aiContext, isGoogleVersion),
+        custom_extensions: customExtensionsDoc(field.customExtensions, `field '${field.name}'`, isGoogleVersion, 
+        /* isModel= */ false, warnings, /* isField= */ true),
+    });
+}
+function metricDoc(metric, warnings, version) {
+    const isGoogleVersion = version === loader_1.GOOGLE_VERSION;
+    const placement = loader_1.FORMAT_VERSION_PLACEMENT[version];
+    // Emit `entity` only from `authoredEntity` (natively under GOOGLE_VERSION or
+    // in the metric's GOOGLE block under OSSIE_VERSION). A derived
+    // `metric.entity` is recomputed by the loader from SQL qualifiers.
+    const expression = expressionDoc(metric, isGoogleVersion, `metric '${metric.name}'`, warnings);
+    if (!expression) {
+        warnings.push(`metric '${metric.name}': no expression; the loader requires one per ` +
+            `metric and the document will not load until it is set.`);
+    }
+    const authoredEntity = metric.authoredEntity || undefined;
+    const blockPayload = placement.metric.entity === 'block' ?
+        compact({ entity: authoredEntity }) :
+        undefined;
+    return compact({
+        name: metric.name,
+        entity: placement.metric.entity === 'native' ? authoredEntity : undefined,
+        expression,
+        datatype: datatypeDoc(metric.type),
+        description: metric.description,
+        ai_context: aiContextDoc(metric.aiContext, isGoogleVersion),
+        custom_extensions: customExtensionsDoc(metric.customExtensions, `metric '${metric.name}'`, isGoogleVersion, 
+        /* isModel= */ false, warnings, /* isField= */ false, blockPayload),
+    });
+}
+// The `Concept.field` pairs this document will still resolve on reload. A
+// projection is only worth emitting as a projection if the field it names
+// comes out in the same document -- see `parameterDoc`. It asks the loader's
+// own resolver rather than walking the model here, so the emitter cannot come
+// to a different answer than the load it is writing for. A resolver that
+// throws leaves every projection emitted as authored: the emitter is not the
+// place to discover that a model does not load.
+function emitted(model) {
+    let concepts;
+    try {
+        concepts = (0, resolve_inheritance_1.declaredConceptFields)(model);
+    }
+    catch {
+        return null;
+    }
+    const pairs = new Set();
+    for (const [name, concept] of concepts) {
+        // A field carrying no emitted datatype resolves to none on reload, so
+        // emitting a projection at it would suppress the parameter's resolved
+        // `type` and hand back an untyped parameter the loader rejects. Treating
+        // the pair as unresolvable falls the parameter back to that resolved type,
+        // which is the whole point of the fallback, and warns when there is none.
+        for (const [field, def] of concept.fields) {
+            if (datatypeDoc(def.type) !== undefined)
+                pairs.add(`${name}.${field}`);
+        }
+    }
+    return pairs;
+}
+// Inverts loader.convertAction. The executor collapses back to the open
+// format's single-key object; parameters emit as their authoring form. Actions
+// only exist under GOOGLE_VERSION, and loader.ts rejects custom_extensions on
+// them.
+function actionDoc(action, warnings, resolvable) {
+    if (action.customExtensions?.length) {
+        warnings.push(`action '${action.name}': ${action.customExtensions
+            .length} custom_extension(s) have no representation ` +
+            `under '${loader_1.GOOGLE_VERSION}' and are not serialized`);
+    }
+    return compact({
+        name: action.name,
+        description: action.description,
+        executor: action.executor ? executorDoc(action.executor) : undefined,
+        parameters: nonEmpty((action.parameters ??
+            []).map(p => parameterDoc(p, resolvable, action.name, warnings))),
+        guards: nonEmpty(action.guards),
+        affects: nonEmpty((action.affects ?? []).map(affectedConceptDoc)),
+        ai_context: aiContextDoc(action.aiContext, /* isGoogleVersion= */ true),
+    });
+}
+// One action parameter, back in the form an author writes.
+//
+// A parameter projected from a field emits the projection and NOT the type.
+// The type is the field's, copied in at load; writing it back out would make
+// the document state a type alongside `concept`/`field`, which is the one
+// thing the loader refuses outright -- so emitting it would produce a document
+// this tool cannot read. The reload resolves it again from the same field and
+// arrives at the same answer.
+//
+// Wording is emitted whether the author wrote it or inherited it, because
+// nothing here can tell those apart and the two reload identically: an
+// authored value that matches the field's is the field's. So this is verbose
+// where the author was terse, and exact either way.
+//
+// `name` is likewise always written, even where the author let it default to
+// the field's name. Same trade, same reason.
+// A projection emits as a projection only where the field it names comes out
+// in this same document. Pruning can take the concept away -- an entity no
+// binding reaches is dropped, and the parameter that projected from it is
+// deliberately kept, because it copied what it needed at load time. Emitting
+// `concept`/`field` anyway would write a document that no longer loads: the
+// pair resolves to nothing, and the type that would have rescued it was
+// suppressed precisely because the projection was supposed to supply it. So
+// where the field is gone, the parameter emits as the declared one it has
+// effectively become -- its resolved `type`, and none of the projection.
+function parameterDoc(p, resolvable, action, warnings) {
+    const projects = p.concept !== undefined &&
+        (resolvable === null || resolvable.has(`${p.concept}.${p.field}`));
+    // The fallback above rests on there BEING a resolved type to fall back to,
+    // and a parameter can reach here without one -- a pull of an aspect written
+    // before the type was recorded, or a model whose inheritance would not
+    // resolve. Both keys then drop and the parameter emits as name and wording
+    // alone, which reloads as a typeless parameter and is rejected by validate.
+    // Nothing here can invent the type, so say so at write time rather than let
+    // it surface as a load error against a document this wrote. `Opaque` is a
+    // valid standalone parameter type in the loader (unlike on fields and
+    // metrics), so `p.type` is emitted directly without `datatypeDoc`.
+    if (!projects && p.concept !== undefined && p.type === undefined) {
+        warnings.push(`action '${action}': parameter '${p.name}' projected from '${p.concept}.${p.field}', which this document does not carry, and ` +
+            `no resolved datatype was recorded to write in its place. The ` +
+            `parameter emits with no type and the document will not load until ` +
+            `one is supplied.`);
+    }
+    return compact({
+        name: p.name,
+        type: projects ? undefined : p.type,
+        concept: projects ? p.concept : undefined,
+        field: projects ? p.field : undefined,
+        label: p.label,
+        description: p.description,
+        ai_context: aiContextDoc(p.aiContext, /* isGoogleVersion= */ true),
+        required: p.required,
+        default: p.default,
+    });
+}
+// One affected concept back to the open format. An entry that says nothing
+// beyond which
+// concept it touches emits as the bare name, which is the shorthand the author
+// most likely wrote and the form the proposal uses; anything with an operation
+// or fields needs the record.
+//
+// This NORMALIZES rather than reproducing the document byte for byte: an
+// authored `{concept: Order}` with no operation comes back as `Order`. The two
+// mean the same thing, so the emit is a fixed point after one pass, which is
+// what the round-trip tests assert.
+function affectedConceptDoc(affected) {
+    if (!affected.operation && !affected.fields?.length)
+        return affected.concept;
+    return compact({
+        concept: affected.concept,
+        operation: affected.operation,
+        fields: nonEmpty(affected.fields),
+    });
+}
+// Inverts loader.convertConstraint. The judgment is a logical invariant and
+// round-trips verbatim, since nothing about it is derived. The aspect type's
+// reserved `expression` and `evaluation` fields have no counterpart here: one
+// names a body the model no longer has, and the other was computed from which
+// body a constraint stated, so emitting either would put something the author
+// cannot write into an authored document.
+function constraintDoc(constraint) {
+    // No custom_extensions call: a constraint carries no custom extensions.
+    return compact({
+        name: constraint.name,
+        judgment: constraint.judgment,
+        description: constraint.description,
+        on_violation: constraint.onViolation,
+        severity: constraint.severity,
+        ai_context: aiContextDoc(constraint.aiContext, /* isGoogleVersion= */ true),
+    });
+}
+// Collapses the IR's tagged executor union back to the open format's single-key
+// object ({mcp: {...}} / {rest: {...}} / {grpc: {...}}).
+function executorDoc(ex) {
+    switch (ex.kind) {
+        case 'mcp':
+            return { mcp: { server: ex.mcp.server, tool: ex.mcp.tool } };
+        case 'rest':
+            return { rest: { endpoint: ex.rest.endpoint, method: ex.rest.method } };
+        case 'grpc':
+            return { grpc: { service: ex.grpc.service, method: ex.grpc.method } };
+        case 'sql':
+            return { sql: { statements: [...ex.sql.statements] } };
+    }
+}
+// Inverts loader.convertRelationship: `from`/`to` are the endpoint entities and
+// `from_columns`/`to_columns` are their positional join columns. An association
+// (junction-table) edge has no open-format syntax, so only this direct-FK view
+// is emitted and the junction detail is flagged.
+function relationshipDoc(rel, warnings, version) {
+    const isGoogleVersion = version === loader_1.GOOGLE_VERSION;
+    const placement = loader_1.FORMAT_VERSION_PLACEMENT[version];
+    if (rel.association) {
+        warnings.push(`relationship '${rel.name}': association (junction-table) detail has no ` +
+            `open-format representation and is not serialized; only its ` +
+            `foreign-key endpoints are written.`);
+    }
+    if (!isGoogleVersion &&
+        (!rel.source.columns.length || !rel.destination.columns.length)) {
+        warnings.push(`relationship '${rel.name}': no join columns; a '${loader_1.OSSIE_VERSION}' document requires from_columns and to_columns ` +
+            `and will not load until they are set.`);
+    }
+    const description = rel.description || undefined;
+    const blockPayload = placement.relationship.description === 'block' ?
+        compact({ description }) :
+        undefined;
+    return compact({
+        name: rel.name,
+        from: rel.source.entity,
+        to: rel.destination.entity,
+        from_columns: nonEmpty(rel.source.columns),
+        to_columns: nonEmpty(rel.destination.columns),
+        description: placement.relationship.description === 'native' ? description :
+            undefined,
+        ai_context: aiContextDoc(rel.aiContext, isGoogleVersion),
+        custom_extensions: customExtensionsDoc(rel.customExtensions, `relationship '${rel.name}'`, isGoogleVersion, 
+        /* isModel= */ false, warnings, /* isField= */ false, blockPayload),
+    });
+}
+// Renders a field or metric `expression` for the target version. The IR comes
+// from two sources:
+//   1. YAML loaded by `loader.ts`, which populates `dialects` and `stringForm`.
+//   2. Models built by `kcmd pull` or OWL import, which leave `dialects` unset
+//      and only populate `expression` (plus optional `importedExpression`).
+function expressionDoc(item, isGoogleVersion, where, warnings) {
+    // Path 1 (YAML round-trip): re-emit the original `dialects` list in authored
+    // order. Under GOOGLE_VERSION, `stringForm` means the author wrote a bare
+    // string shorthand (`expression: "SUM(amt)"`), so emit the bare string back;
+    // under OSSIE_VERSION bare strings are forbidden, so emit `{dialects:
+    // [...]}`.
+    if (item.dialects?.length) {
+        const allowed = [];
+        for (const d of item.dialects) {
+            if (ALLOWED_DIALECT_SET.has(d.dialect)) {
+                allowed.push(d);
+            }
+            else {
+                warnings.push(`${where}: expression in dialect '${d.dialect}' is not a dialect ` +
+                    `the loader accepts and is not serialized`);
+            }
+        }
+        if (allowed.length) {
+            if (isGoogleVersion && item.stringForm) {
+                return allowed[0].expression;
+            }
+            return {
+                dialects: allowed.map(d => ({ dialect: d.dialect, expression: d.expression })),
+            };
+        }
+    }
+    // Path 2 (`kcmd pull` / OWL import / fallback): `dialects` is absent, so
+    // synthesize it from `expression` and `importedExpression`.
+    if (isGoogleVersion && item.stringForm && item.expression !== undefined &&
+        item.importedExpression === undefined) {
+        return item.expression;
+    }
+    const dialects = [];
+    if (item.expression !== undefined) {
+        // `loader.ts` wraps a bare-string expression as ANSI_SQL; otherwise label
+        // the canonical expression BIGQUERY so reload picks it back up.
+        dialects.push({
+            dialect: item.stringForm ? 'ANSI_SQL' : CANONICAL_DIALECT,
+            expression: item.expression,
+        });
+    }
+    if (item.importedExpression !== undefined) {
+        // `importedExpression` holds third-party vendor SQL, such as SNOWFLAKE or
+        // DATABRICKS, kept for transpilation. Only emit `importedDialect` when the
+        // loader accepts it; an unknown or missing label would fail to reload.
+        const label = item.importedDialect;
+        if (label === undefined) {
+            warnings.push(`${where}: imported expression has no dialect and is not serialized`);
+        }
+        else if (!ALLOWED_DIALECT_SET.has(label)) {
+            warnings.push(`${where}: imported expression in dialect '${label}' is not a ` +
+                `dialect the loader accepts and is not serialized`);
+        }
+        else if (dialects.some(d => d.dialect === label)) {
+            warnings.push(`${where}: imported expression in dialect '${label}' has the same ` +
+                `dialect as the expression and is not serialized`);
+        }
+        else {
+            dialects.push({ dialect: label, expression: item.importedExpression });
+        }
+    }
+    return dialects.length ? { dialects } : undefined;
+}
+// Emits the field's `dimension` block when present, inverting
+// loader.convertField (which sets `dimension = {}` for a bare marker and copies
+// `is_time`). The key is emitted whenever the IR carries dimension metadata,
+// even for an empty marker, so a dimension field reloads as a dimension.
+function dimensionDoc(field) {
+    if (!field.dimension)
+        return undefined;
+    return compact({ is_time: field.dimension.isTime });
+}
+// Emits the structured `ai_context` for the target version, inverting
+// loader.normalizeAiContext. Custom properties (`additionalProperties`) sit
+// under `custom:` under GOOGLE_VERSION and as sibling keys under OSSIE_VERSION.
+function aiContextDoc(ai, isGoogleVersion) {
+    if (!ai)
+        return undefined;
+    const custom = ai.additionalProperties ? compact(ai.additionalProperties) : undefined;
+    const hasCustom = custom !== undefined && Object.keys(custom).length > 0;
+    const base = compact({
+        instructions: ai.instructions,
+        synonyms: nonEmpty(ai.synonyms),
+        examples: nonEmpty(ai.examples),
+    });
+    if (!hasCustom) {
+        return Object.keys(base).length ? base : undefined;
+    }
+    return isGoogleVersion ? { ...base, custom } : { ...custom, ...base };
+}
+// The deployments a vanilla model lists only as targets in its GOOGLE blocks,
+// named as a pull names them. A block whose data is not valid JSON gives none,
+// and customExtensionsDoc warns that it was dropped. The other blocks still
+// give theirs.
+function deploymentsFromGoogleExtensions(model) {
+    const targets = (model.customExtensions ??
+        []).flatMap(ext => blockTargets(ext, model.name));
+    return (0, deployment_target_1.unnamedDeployments)(targets);
+}
+// Serializes `custom_extensions` for one IR element. Under OSSIE_VERSION,
+// third-party extensions are emitted, followed by a synthesized GOOGLE block
+// when `blockPayload` carries any keys. Any other raw GOOGLE block on the IR
+// is dropped with a warning so it cannot collide with the synthesized block or
+// write keys the loader rejects. Under GOOGLE_VERSION,
+// third-party extensions are written out unchanged without warning, while
+// `GOOGLE` blocks are skipped (silently for the model's `deploymentTargets`
+// block, and with a warning for any other `GOOGLE` block).
+function customExtensionsDoc(exts, where, isGoogleVersion, isModel, warnings, isField = false, blockPayload) {
+    const payload = blockPayload ? { ...blockPayload } : {};
+    if (!exts?.length && Object.keys(payload).length === 0)
+        return undefined;
+    if (!isGoogleVersion) {
+        const kept = [];
+        let droppedGoogle = 0;
+        for (const e of exts ?? []) {
+            if (e.vendorName !== GOOGLE_VENDOR) {
+                kept.push({ vendor_name: e.vendorName, data: e.data });
+                continue;
+            }
+            // The loader's copy of the model's `deployments`, which the GOOGLE
+            // block below writes in their own form.
+            if (isModel && payload.deployments !== undefined &&
+                isDeploymentTargetsBlock(e)) {
+                continue;
+            }
+            droppedGoogle += 1;
+        }
+        if (droppedGoogle > 0) {
+            warnings.push(isField ?
+                `${where}: ${droppedGoogle} GOOGLE custom_extension(s) have no ` +
+                    `representation on a field under '${loader_1.OSSIE_VERSION}' and ` +
+                    `are not serialized` :
+                `${where}: ${droppedGoogle} GOOGLE custom_extension(s) are not ` +
+                    `serialized under '${loader_1.OSSIE_VERSION}'; its GOOGLE block is ` +
+                    `written from its own fields`);
+        }
+        if (Object.keys(payload).length > 0) {
+            kept.push({
+                vendor_name: GOOGLE_VENDOR,
+                data: JSON.stringify(payload),
+            });
+        }
+        return nonEmpty(kept);
+    }
+    // No GOOGLE_VERSION construct has 'block' placement, so `blockPayload` is
+    // empty here.
+    const kept = [];
+    let droppedGoogle = 0;
+    for (const ext of exts ?? []) {
+        if (ext.vendorName !== GOOGLE_VENDOR) {
+            kept.push({ vendor_name: ext.vendorName, data: ext.data });
+            continue;
+        }
+        if (isModel && blockTargets(ext, '').length > 0)
+            continue;
+        droppedGoogle += 1;
+    }
+    if (droppedGoogle > 0) {
+        warnings.push(`${where}: ${droppedGoogle} GOOGLE custom_extension(s) have no ` +
+            `representation under '${loader_1.GOOGLE_VERSION}' and are not serialized`);
+    }
+    return nonEmpty(kept);
+}
+/**
+ * Returns true if `e.data` is a JSON object with a `'deploymentTargets'`
+ * property.
+ *
+ * The loader folds a Google-flavor model's `deployments` into such a block.
+ */
+function isDeploymentTargetsBlock(e) {
+    try {
+        const d = JSON.parse(e.data);
+        return !!d && typeof d === 'object' &&
+            Object.hasOwn(d, 'deploymentTargets');
+    }
+    catch {
+        return false;
+    }
+}
+function blockTargets(ext, modelName) {
+    try {
+        return (0, deployment_target_1.googleBlockTargets)([ext], modelName);
+    }
+    catch {
+        return [];
+    }
+}
+// Drops undefined-valued keys so the emitted YAML only shows fields the model
+// actually set (matching the emitters' `compact` convention).
+function compact(obj) {
+    const out = {};
+    for (const [k, v] of Object.entries(obj)) {
+        if (v !== undefined)
+            out[k] = v;
+    }
+    return out;
+}
+// Returns the array, or undefined when empty/absent, so an empty list is
+// omitted rather than rendered as `[]`.
+function nonEmpty(items) {
+    return items && items.length ? items : undefined;
+}