app/api/endpoints/household.py

@@ -11,8 +11,7 @@ import sqlite3
 import requests
 from fastapi import APIRouter, Depends, HTTPException
 
-from app.cloud_session import CloudUser, CLOUD_DATA_ROOT, get_session
+from app.cloud_session import CloudUser, CLOUD_DATA_ROOT, HEIMDALL_URL, HEIMDALL_ADMIN_TOKEN, get_session
-from app.services.heimdall_orch import HEIMDALL_URL, HEIMDALL_ADMIN_TOKEN
 from app.db.store import Store
 from app.models.schemas.household import (
     HouseholdAcceptRequest,

app/api/endpoints/saved_recipes.py

@@ -35,7 +35,7 @@ def _to_summary(row: dict, store: Store) -> SavedRecipeSummary:
     return SavedRecipeSummary(
         id=row["id"],
         recipe_id=row["recipe_id"],
-        title=row.get("title") or "",
+        title=row.get("title", ""),
         saved_at=row["saved_at"],
         notes=row.get("notes"),
         rating=row.get("rating"),
@@ -104,10 +104,8 @@ async def list_saved_recipes(
 async def list_collections(
     session: CloudUser = Depends(get_session),
 ) -> list[CollectionSummary]:
-    # Free users can list (they'll always have zero — creating requires Paid).
-    # Returning 403 here breaks savedStore.load() via Promise.all for non-Paid users.
     if not can_use("recipe_collections", session.tier):
-        return []
+        raise HTTPException(status_code=403, detail="Collections require Paid tier.")
     rows = await asyncio.to_thread(
         _in_thread, session.db, lambda s: s.get_collections()
     )

app/db/migrations/039_saved_recipes_drop_fk.sql

@@ -1,31 +0,0 @@
--- Migration 039: Drop FK constraint on saved_recipes.recipe_id.
---
--- In cloud mode the recipe corpus is ATTACHed as a separate database.
--- SQLite FK constraints only resolve against the `main` schema, so
--- `REFERENCES recipes(id)` was always failing for cloud saves (the
--- main.recipes table is empty; all data lives in corpus.recipes).
--- The corpus is read-only and never modified by the app, so cascade-on-delete
--- is meaningless anyway. Remove the constraint without changing any data.
-
-PRAGMA foreign_keys = OFF;
-
-CREATE TABLE saved_recipes_new (
-    id          INTEGER PRIMARY KEY AUTOINCREMENT,
-    recipe_id   INTEGER NOT NULL,
-    saved_at    TEXT    NOT NULL DEFAULT (datetime('now')),
-    notes       TEXT,
-    rating      INTEGER CHECK (rating IS NULL OR (rating >= 0 AND rating <= 5)),
-    style_tags  TEXT    NOT NULL DEFAULT '[]',
-    UNIQUE (recipe_id)
-);
-
-INSERT INTO saved_recipes_new SELECT * FROM saved_recipes;
-
-DROP TABLE saved_recipes;
-
-ALTER TABLE saved_recipes_new RENAME TO saved_recipes;
-
-CREATE INDEX IF NOT EXISTS idx_saved_recipes_saved_at ON saved_recipes (saved_at DESC);
-CREATE INDEX IF NOT EXISTS idx_saved_recipes_rating   ON saved_recipes (rating);
-
-PRAGMA foreign_keys = ON;

app/models/schemas/recipe.py

@@ -4,27 +4,6 @@ from __future__ import annotations
 from pydantic import BaseModel, Field
 
 
-class StepAnalysis(BaseModel):
-    """Active/passive classification for one direction step."""
-    is_passive: bool
-    detected_minutes: int | None = None
-
-
-class TimeEffortProfile(BaseModel):
-    """Parsed time and effort profile for a recipe.
-
-    Mirrors app.services.recipe.time_effort.TimeEffortProfile (dataclass).
-    Serialised into RecipeSuggestion so the frontend can render the effort
-    summary without a second round-trip.
-    """
-    active_min: int = 0
-    passive_min: int = 0
-    total_min: int = 0
-    effort_label: str = "moderate"   # "quick" | "moderate" | "involved"
-    equipment: list[str] = Field(default_factory=list)
-    step_analyses: list[StepAnalysis] = Field(default_factory=list)
-
-
 class SwapCandidate(BaseModel):
     original_name: str
     substitute_name: str
@@ -64,7 +43,6 @@ class RecipeSuggestion(BaseModel):
     source_url: str | None = None
     complexity: str | None = None  # 'easy' | 'moderate' | 'involved'
     estimated_time_min: int | None = None  # derived from step count + method signals
-    time_effort: TimeEffortProfile | None = None  # full time/effort profile from parse_time_effort
     rerank_score: float | None = None  # cross-encoder relevance score (paid+ only, None for free tier)
 
 

app/services/recipe/recipe_engine.py

@@ -20,7 +20,7 @@ from typing import TYPE_CHECKING
 if TYPE_CHECKING:
     from app.db.store import Store
 
-from app.models.schemas.recipe import GroceryLink, NutritionPanel, RecipeRequest, RecipeResult, RecipeSuggestion, StepAnalysis, TimeEffortProfile, SwapCandidate
+from app.models.schemas.recipe import GroceryLink, NutritionPanel, RecipeRequest, RecipeResult, RecipeSuggestion, SwapCandidate
 from app.services.recipe.element_classifier import ElementClassifier
 from app.services.recipe.grocery_links import GroceryLinkBuilder
 from app.services.recipe.substitution_engine import SubstitutionEngine
@@ -36,38 +36,6 @@ _SWAP_STOPWORDS = frozenset({
     "to", "from", "at", "by", "as", "on",
 })
 
-# Marketing / prep / packaging words stripped when tokenising product-label names
-# into individual ingredient tokens.  Parallel to Store._FTS_TOKEN_STOPWORDS —
-# both lists should agree.  Kept here to avoid a circular import at runtime.
-_PRODUCT_TOKEN_STOPWORDS = frozenset({
-    # Basic English stopwords
-    "a", "an", "the", "of", "in", "for", "with", "and", "or", "to",
-    "from", "at", "by", "as", "on", "into",
-    # Brand / marketing words that appear in product names
-    "lean", "cuisine", "healthy", "choice", "stouffer", "original",
-    "classic", "deluxe", "homestyle", "family", "style", "grade",
-    "premium", "select", "natural", "organic", "fresh", "lite",
-    "ready", "quick", "easy", "instant", "microwave", "frozen",
-    "brand", "size", "large", "small", "medium", "extra",
-    # Plant-based / alt-meat brand names
-    "daring", "gardein", "morningstar", "lightlife", "tofurky",
-    "quorn", "omni", "nuggs", "simulate",
-    # Preparation states
-    "cut", "diced", "sliced", "chopped", "minced", "shredded",
-    "cooked", "raw", "whole", "boneless", "skinless", "trimmed",
-    "pre", "prepared", "marinated", "seasoned", "breaded", "battered",
-    "grilled", "roasted", "smoked", "canned", "dried", "dehydrated",
-    "pieces", "piece", "strips", "strip", "chunks", "chunk",
-    "fillets", "fillet", "cutlets", "cutlet", "tenders", "nuggets",
-    # Units / packaging
-    "oz", "lb", "lbs", "pkg", "pack", "box", "can", "bag", "jar",
-    # Adjectives that aren't ingredients
-    "firm", "soft", "silken", "hard", "crispy", "crunchy", "smooth",
-    "mild", "spicy", "hot", "sweet", "savory", "unsalted", "salted",
-    "low", "high", "reduced", "free", "fat", "sodium", "sugar", "calorie",
-    "dairy", "gluten", "vegan", "plant", "based", "free",
-})
-
 # Maps product-label substrings to recipe-corpus canonical terms.
 # Kept in sync with Store._FTS_SYNONYMS — both must agree on canonical names.
 # Used to expand pantry_set so single-word recipe ingredients can match
@@ -395,13 +363,6 @@ def _expand_pantry_set(
             if pattern in lower:
                 expanded.add(canonical)
 
-        # Extract individual ingredient tokens from multi-word product names.
-        # "Organic Extra Firm Tofu" → adds "tofu"; "Brown Basmati Rice" → adds "rice".
-        # This catches plain ingredients that _PANTRY_LABEL_SYNONYMS doesn't translate.
-        for token in lower.split():
-            if len(token) >= 4 and token not in _PRODUCT_TOKEN_STOPWORDS:
-                expanded.add(token)
-
         # Secondary state expansion — adds terms like "stale bread", "day-old rice"
         if secondary_pantry_items and item in secondary_pantry_items:
             state_label = secondary_pantry_items[item]
@@ -775,13 +736,9 @@ class RecipeEngine:
         #   - match ratio: require ≥60% ingredient coverage to avoid low-signal results
         _l1 = req.level == 1 and not req.shopping_mode
         nf = req.nutrition_filters
-        # L1 uses a larger candidate pool — the ratio gate below will prune
-        # aggressively anyway, so we need more raw candidates to end up with
-        # enough results for a packaged-food / plant-based pantry.
-        _fts_limit = 60 if _l1 else 20
         rows = self._store.search_recipes_by_ingredients(
             req.pantry_items,
-            limit=_fts_limit,
+            limit=20,
             category=req.category or None,
             max_calories=nf.max_calories,
             max_sugar_g=nf.max_sugar_g,
@@ -792,11 +749,8 @@ class RecipeEngine:
         )
 
         # L1 strict defaults: cap missing ingredients and require a minimum ratio.
-        # 0.35 allows ~1/3 ingredient coverage — low enough for packaged/plant-based
-        # pantries that rarely match raw-ingredient corpus recipes 1:1, but still
-        # filters out recipes where only one common staple matched.
         _L1_MAX_MISSING_DEFAULT = 2
-        _L1_MIN_MATCH_RATIO = 0.35
+        _L1_MIN_MATCH_RATIO = 0.6
         effective_max_missing = req.max_missing
         if _l1 and effective_max_missing is None:
             effective_max_missing = _L1_MAX_MISSING_DEFAULT
@@ -880,10 +834,9 @@ class RecipeEngine:
                 except Exception:
                     directions = [directions]
 
-            # Compute complexity + parse time effort once — reused for filters and response.
+            # Compute complexity for every suggestion (used for badge + filter).
             row_complexity = _classify_method_complexity(directions, available_equipment)
             row_time_min = _estimate_time_min(directions, row_complexity)
-            row_time_effort = parse_time_effort(directions)
 
             # Filter and tier-rank by hard_day_mode
             if req.hard_day_mode:
@@ -903,15 +856,8 @@ class RecipeEngine:
             if req.max_time_min is not None and row_time_min > req.max_time_min:
                 continue
 
-            # Total time filter (kiwi#52).
+            # Total time filter (kiwi#52) — uses parsed time from directions
-            # Prefer parsed time extracted from direction text (explicit "15 minutes" mentions).
+            if req.max_total_min is not None and not _within_time(directions, req.max_total_min):
-            # When directions contain no parseable time signals, fall back to the
-            # step-count estimate so the filter still has teeth on the corpus majority.
-            if req.max_total_min is not None:
-                if row_time_effort.total_min > 0:
-                    if row_time_effort.total_min > req.max_total_min:
-                        continue
-                elif row_time_min > req.max_total_min:
                 continue
 
             # Level 2: also add dietary constraint swaps from substitution_pairs
@@ -951,20 +897,6 @@ class RecipeEngine:
                 v is not None
                 for v in (nutrition.calories, nutrition.sugar_g, nutrition.carbs_g)
             )
-            te = TimeEffortProfile(
-                active_min=row_time_effort.active_min,
-                passive_min=row_time_effort.passive_min,
-                total_min=row_time_effort.total_min,
-                effort_label=row_time_effort.effort_label,
-                equipment=list(row_time_effort.equipment),
-                step_analyses=[
-                    StepAnalysis(
-                        is_passive=sa.is_passive,
-                        detected_minutes=sa.detected_minutes,
-                    )
-                    for sa in row_time_effort.step_analyses
-                ],
-            )
             suggestions.append(RecipeSuggestion(
                 id=row["id"],
                 title=row["title"],
@@ -973,14 +905,12 @@ class RecipeEngine:
                 swap_candidates=swap_candidates,
                 matched_ingredients=matched,
                 missing_ingredients=missing,
-                directions=directions,
                 prep_notes=sorted(prep_note_set),
                 level=req.level,
                 nutrition=nutrition if has_nutrition else None,
                 source_url=_build_source_url(row),
                 complexity=row_complexity,
                 estimated_time_min=row_time_min,
-                time_effort=te,
             ))
 
         # Sort corpus results.

frontend/src/components/RecipesView.vue

@@ -513,12 +513,12 @@
               <span v-if="recipe.estimated_time_min" class="status-badge status-neutral">~{{ recipe.estimated_time_min }}m</span>
               <span class="status-badge status-info">Level {{ recipe.level }}</span>
               <span v-if="recipe.is_wildcard" class="status-badge status-info">Wildcard</span>
-              <span
+              <button
                 v-if="recipe.id"
-                :class="['btn-icon', 'btn-bookmark', { active: savedStore.isSaved(recipe.id) }]"
+                :class="['btn-icon', 'btn-bookmark', { active: recipesStore.isBookmarked(recipe.id) }]"
-                :aria-label="savedStore.isSaved(recipe.id) ? 'Saved: ' + recipe.title : recipe.title"
+                @click="recipesStore.toggleBookmark(recipe)"
-                :title="savedStore.isSaved(recipe.id) ? 'Saved' : 'Not saved'"
+                :aria-label="recipesStore.isBookmarked(recipe.id) ? 'Remove bookmark: ' + recipe.title : 'Bookmark: ' + recipe.title"
-              >{{ savedStore.isSaved(recipe.id) ? '★' : '☆' }}</span>
+              >{{ recipesStore.isBookmarked(recipe.id) ? '★' : '☆' }}</button>
               <button
                 v-if="recipe.id"
                 class="btn-icon btn-dismiss"
@@ -1212,9 +1212,7 @@ onMounted(async () => {
   await savedStore.load()
 })
 
-// If Saved tab is empty after loading, bounce to Build Your Own.
+// If Saved tab is empty after loading, bounce to Build Your Own
-// No immediate: true — the immediate fire happens before onMounted runs load(),
-// so loading=false and count=0 is the initial unloaded state, not "empty after load".
 watch(
   () => ({ loading: savedStore.loading, count: savedStore.saved.length }),
   ({ loading, count }) => {
@@ -1222,6 +1220,7 @@ watch(
       activeTab.value = 'build'
     }
   },
+  { immediate: true },
 )
 </script>
 

frontend/src/stores/savedRecipes.ts

@@ -27,15 +27,12 @@ export const useSavedRecipesStore = defineStore('savedRecipes', () => {
   async function load() {
     loading.value = true
     try {
-      // Fetch independently — a collections 403 (Free tier) must not prevent
+      const [items, cols] = await Promise.all([
-      // saved recipes from loading. Backend now returns [] for Free, but guard
-      // here too in case an older API version is deployed.
-      const [itemsResult, colsResult] = await Promise.allSettled([
         savedRecipesAPI.list({ sort_by: sortBy.value, collection_id: activeCollectionId.value ?? undefined }),
         savedRecipesAPI.listCollections(),
       ])
-      if (itemsResult.status === 'fulfilled') saved.value = itemsResult.value
+      saved.value = items
-      if (colsResult.status === 'fulfilled') collections.value = colsResult.value
+      collections.value = cols
     } finally {
       loading.value = false
     }

tests/fixtures/recipe_scan/extract_test.py

@@ -1,107 +0,0 @@
-#!/usr/bin/env python3
-"""
-Prompt validation harness for recipe scanner (kiwi#9).
-
-Runs the draft extraction prompt against fixture images using the Anthropic API
-directly (bypasses llm.yaml — for prompt dev only, not production path).
-
-Usage:
-    python extract_test.py <image1.jpg> [image2.jpg]
-"""
-import base64
-import io
-import json
-import os
-import sys
-from pathlib import Path
-
-from PIL import Image, ImageOps
-import anthropic
-
-PROMPT = """
-You are extracting a recipe from a photograph of a recipe card, cookbook page, or handwritten note.
-
-If two images are provided, treat them as a single recipe across two pages (e.g. ingredients on page 1, directions on page 2).
-
-Return a single JSON object with these fields:
-- title: recipe name (string)
-- subtitle: any secondary title or serving suggestion e.g. "with Broccoli & Ranch Dressing" (string or null)
-- servings: serving size if shown, as a string e.g. "2", "4-6" (string or null)
-- cook_time: total cook time if shown, e.g. "15 min", "1 hour" (string or null)
-- source_note: any attribution text like "From Betty Crocker" or "Purple Carrot" (string or null)
-- ingredients: array of ingredient objects, each with:
-  - name: normalized generic ingredient name, lowercase, no quantities, no brand names
-    (e.g. "Follow Your Heart® Vegan Ranch" → "ranch dressing")
-  - qty: quantity as a string, preserving fractions e.g. "1/2", "¼" (string or null)
-  - unit: unit of measure, null for countable items (e.g. "3 eggs" → unit: null)
-  - raw: the original ingredient line verbatim, exactly as it appears
-- steps: ordered array of instruction strings, one distinct step per element
-- notes: any tips, substitutions, storage instructions, or variations (string or null)
-- confidence: "high" if text is clear and complete, "medium" if some parts are uncertain,
-  "low" if mostly handwritten or significantly degraded
-- warnings: array of strings describing anything the user should double-check
-  (e.g. "Directions appear to continue on another page not shown")
-
-Return only valid JSON. No markdown fences. No explanation outside the JSON.
-If the image does not appear to be a recipe at all, return: {"error": "not_a_recipe"}
-""".strip()
-
-
-def load_image_b64(path: Path) -> str:
-    """Load image, apply EXIF rotation, return base64-encoded JPEG."""
-    with open(path, "rb") as f:
-        img = Image.open(io.BytesIO(f.read()))
-        img = ImageOps.exif_transpose(img)  # fix phone rotation
-        img = img.convert("RGB")
-        buf = io.BytesIO()
-        img.save(buf, format="JPEG", quality=90)
-        return base64.b64encode(buf.getvalue()).decode()
-
-
-def extract(image_paths: list[Path]) -> dict:
-    client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])
-
-    content = []
-    for i, path in enumerate(image_paths):
-        if i > 0:
-            content.append({"type": "text", "text": f"(Page {i + 1} of the same recipe:)"})
-        content.append({
-            "type": "image",
-            "source": {
-                "type": "base64",
-                "media_type": "image/jpeg",
-                "data": load_image_b64(path),
-            },
-        })
-    content.append({"type": "text", "text": PROMPT})
-
-    msg = client.messages.create(
-        model="claude-opus-4-6",  # best vision for prompt dev; production uses VisionRouter
-        max_tokens=2048,
-        messages=[{"role": "user", "content": content}],
-    )
-    raw = msg.content[0].text.strip()
-    # Strip markdown fences if the model adds them anyway
-    if raw.startswith("```"):
-        raw = raw.split("```")[1]
-        if raw.startswith("json"):
-            raw = raw[4:]
-    return json.loads(raw)
-
-
-if __name__ == "__main__":
-    paths = [Path(p) for p in sys.argv[1:]]
-    if not paths:
-        print("Usage: python extract_test.py <image1.jpg> [image2.jpg]")
-        sys.exit(1)
-
-    for p in paths:
-        if not p.exists():
-            print(f"File not found: {p}")
-            sys.exit(1)
-
-    print(f"Extracting from: {[p.name for p in paths]}")
-    print("Applying EXIF rotation + sending to claude-opus-4-6...\n")
-
-    result = extract(paths)
-    print(json.dumps(result, indent=2, ensure_ascii=False))