feat(config): add GPU_SERVER_URL alias for CF_ORCH_URL

Self-hoster-friendly env var name. Priority: GPU_SERVER_URL →
CF_ORCH_URL (compat) → https://orch.circuitforge.tech when
CF_LICENSE_KEY is present (Paid+ auto-default). Resolved value
written back to os.environ["CF_ORCH_URL"] at startup so all
service callers remain unchanged.

Bump version to 0.10.0.

.env.example

@@ -21,10 +21,12 @@ DATA_DIR=./data
 # IP this machine advertises to the coordinator (must be reachable from coordinator host)
 # CF_ORCH_ADVERTISE_HOST=10.1.10.71
 
-# CF-core hosted coordinator (managed cloud GPU inference — Paid+ tier)
-# Set CF_ORCH_URL to use a hosted cf-orch coordinator instead of self-hosting.
-# CF_LICENSE_KEY is read automatically by CFOrchClient for bearer auth.
-# CF_ORCH_URL=https://orch.circuitforge.tech
+# GPU inference server (cf-orch coordinator for recipe scan, LLM generation, etc.)
+# GPU_SERVER_URL: set to your local cf-orch coordinator (self-hosted rack).
+#   CF_ORCH_URL is the backward-compat alias — both are honoured.
+# Paid+ default: when CF_LICENSE_KEY is present and neither URL is set,
+#   the app automatically points to https://orch.circuitforge.tech.
+# GPU_SERVER_URL=http://10.1.10.71:7700
 # CF_LICENSE_KEY=CFG-KIWI-xxxx-xxxx-xxxx
 
 # LLM backend — env-var auto-config (no llm.yaml needed for bare-metal users)
@@ -57,6 +59,9 @@ CF_APP_NAME=kiwi
 # Unset = auto-detect: true if CLOUD_MODE or circuitforge_orch is installed (paid+ local).
 # Set false to force LocalScheduler even when cf-orch is present.
 # USE_ORCH_SCHEDULER=false
+# GPU_SERVER_URL: cf-orch coordinator endpoint. Required for recipe scan (cf-docuvision)
+# and LLM features on a self-hosted rack. CF_ORCH_URL is the backward-compat alias.
+# GPU_SERVER_URL=http://10.1.10.71:7700
 
 # Cloud mode (set in compose.cloud.yml; also set here for reference)
 # CLOUD_DATA_ROOT=/devl/kiwi-cloud-data

app/core/config.py

@@ -65,9 +65,24 @@ class Settings:
     # Quality
     MIN_QUALITY_SCORE: float = float(os.environ.get("MIN_QUALITY_SCORE", "50.0"))
 
-    # CF-core resource coordinator (VRAM lease management)
+    # CF-core resource coordinator (VRAM lease management — lease broker, not inference)
     COORDINATOR_URL: str = os.environ.get("COORDINATOR_URL", "http://localhost:7700")
 
+    # GPU inference server URL
+    # Priority: GPU_SERVER_URL env var → CF_ORCH_URL env var (backward compat)
+    #           → https://orch.circuitforge.tech when CF_LICENSE_KEY is present (Paid+)
+    # Resolved value is written back to os.environ["CF_ORCH_URL"] at startup so
+    # all service-layer callers that read CF_ORCH_URL directly see the right URL.
+    GPU_SERVER_URL: str | None = (
+        os.environ.get("GPU_SERVER_URL")
+        or os.environ.get("CF_ORCH_URL")
+        or (
+            "https://orch.circuitforge.tech"
+            if os.environ.get("CF_LICENSE_KEY")
+            else None
+        )
+    )
+
     # Hosted cf-orch coordinator — bearer token for managed cloud GPU inference (Paid+)
     # CFOrchClient reads CF_LICENSE_KEY automatically; exposed here for startup validation.
     CF_LICENSE_KEY: str | None = os.environ.get("CF_LICENSE_KEY")
@@ -108,3 +123,9 @@ class Settings:
 
 
 settings = Settings()
+
+# Normalise GPU_SERVER_URL into CF_ORCH_URL so every service-layer caller that
+# reads os.environ.get("CF_ORCH_URL") sees the resolved value, including the
+# Paid+ cloud default injected above.
+if settings.GPU_SERVER_URL:
+    os.environ["CF_ORCH_URL"] = settings.GPU_SERVER_URL

app/services/recipe/recipe_scanner.py

@@ -215,6 +215,35 @@ def _build_ocr_extraction_prompt(ocr_text: str) -> str:
     )
 
 
+def _call_via_cf_text_vlm(alloc_url: str, image_paths: list[Path], prompt: str) -> str:
+    """Call the cf-text OpenAI-compat API with images via the llama.cpp multimodal backend."""
+    import httpx
+
+    content: list[dict] = []
+    for i, path in enumerate(image_paths):
+        if i > 0:
+            content.append({"type": "text", "text": f"(Page {i + 1} of the same recipe:)"})
+        b64 = _load_image_b64(path)
+        content.append({
+            "type": "image_url",
+            "image_url": {"url": f"data:image/jpeg;base64,{b64}"},
+        })
+    content.append({"type": "text", "text": prompt})
+
+    resp = httpx.post(
+        f"{alloc_url.rstrip('/')}/v1/chat/completions",
+        json={
+            "model": "local",
+            "messages": [{"role": "user", "content": content}],
+            "max_tokens": 2048,
+            "temperature": 0.0,
+        },
+        timeout=180.0,
+    )
+    resp.raise_for_status()
+    return resp.json()["choices"][0]["message"]["content"].strip()
+
+
 def _call_vision_backend(
     image_paths: list[Path],
     prompt: str,
@@ -222,7 +251,7 @@ def _call_vision_backend(
 ) -> str:
     """Dispatch to the best available vision backend.
 
-    Priority: cf-orch docuvision (OCR + text LLM) -> local Qwen2.5-VL -> Anthropic API.
+    Priority: cf-orch (Qwen2-VL GGUF via cf-text) -> local Qwen2.5-VL -> Anthropic API.
     Raises RuntimeError with a clear message when no backend is available.
 
     Args:
@@ -237,9 +266,8 @@ def _call_vision_backend(
 
     errors: list[str] = []
 
-    # 1. Try cf-orch task allocation → cf-docuvision OCR, then text LLM structuring.
-    #    Two-step: docuvision extracts text from the image(s), then LLMRouter
-    #    converts the OCR text to structured recipe JSON using the extraction prompt.
+    # 1. Try cf-orch task allocation → cf-docuvision (Qwen2-VL GGUF via llama.cpp).
+    #    Two-step: docuvision OCRs the image(s), then LLMRouter structures the text into JSON.
     cf_orch_url = os.environ.get("CF_ORCH_URL")
     if cf_orch_url:
         try:
@@ -250,7 +278,6 @@ def _call_vision_backend(
             try:
                 _progress("allocating", "Starting vision service...")
                 with task_allocate("kiwi", "recipe_scan", service_hint="cf-docuvision", ttl_s=120.0) as alloc:
-                    # Step 1: OCR each image via cf-docuvision
                     _progress("scanning", "Extracting recipe text from photo...")
                     doc_client = DocuvisionClient(alloc.url)
                     ocr_parts: list[str] = []
@@ -263,9 +290,11 @@ def _call_vision_backend(
                     if not combined_ocr.strip():
                         raise ValueError("Docuvision returned no text — image may not be a recipe")
 
-                    # Step 2: Text LLM structures OCR output into recipe JSON
                     _progress("structuring", "Parsing recipe structure...")
-                    text = LLMRouter().complete(_build_ocr_extraction_prompt(combined_ocr))
+                    text = LLMRouter().complete(
+                        _build_ocr_extraction_prompt(combined_ocr),
+                        system="You are a recipe data extractor. Return ONLY valid JSON. No markdown, no explanation, no code fences.",
+                    )
                     if text:
                         return text
 
@@ -303,40 +332,76 @@ def _normalize_ingredient_name(name: str) -> str:
     return name.lower().strip()
 
 
+def _extract_json_object(text: str) -> str | None:
+    """Return the first balanced JSON object from text, or None if not found.
+
+    Uses brace-counting rather than a greedy regex so trailing prose and
+    nested objects are handled correctly.
+    """
+    start = text.find("{")
+    if start == -1:
+        return None
+    depth = 0
+    in_string = False
+    escape_next = False
+    for i, ch in enumerate(text[start:], start):
+        if escape_next:
+            escape_next = False
+            continue
+        if ch == "\\" and in_string:
+            escape_next = True
+            continue
+        if ch == '"':
+            in_string = not in_string
+            continue
+        if in_string:
+            continue
+        if ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return text[start : i + 1]
+    return None
+
+
 def _parse_scanner_json(raw_text: str) -> dict:
     """Extract and return the JSON dict from VLM output.
 
     Handles:
     - Pure JSON
-    - JSON wrapped in ```json ... ``` markdown fences
-    - JSON preceded by a line of prose ("Here is the recipe: {...}")
+    - JSON in ```json ... ``` markdown fences
+    - Qwen3-style <think>...</think> or <thinking>...</thinking> preambles
+    - JSON preceded or followed by prose
 
     Raises ValueError on not_a_recipe or unparseable output.
     """
     text = raw_text.strip()
 
-    # Strip markdown fences if present
-    if text.startswith("```"):
-        parts = text.split("```")
-        for part in parts:
-            part = part.strip()
-            if part.startswith("json"):
-                part = part[4:].strip()
-            if part.startswith("{"):
-                text = part
-                break
+    # Strip thinking-token blocks emitted by reasoning models (Qwen3, DeepSeek-R1, etc.)
+    text = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL | re.IGNORECASE).strip()
+    text = re.sub(r"<thinking>.*?</thinking>", "", text, flags=re.DOTALL | re.IGNORECASE).strip()
 
-    # Try direct parse first
+    # Strip markdown fences if present
+    if "```" in text:
+        # Find the content between the first ``` pair
+        fence_match = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, re.DOTALL)
+        if fence_match:
+            text = fence_match.group(1).strip()
+
+    # Try direct parse
     try:
         data = json.loads(text)
     except json.JSONDecodeError:
-        # Extract first JSON object embedded in prose
-        match = re.search(r"\{.*\}", text, re.DOTALL)
-        if not match:
+        # Fall back to brace-balanced extraction from anywhere in the output
+        candidate = _extract_json_object(text)
+        if not candidate:
+            logger.warning("Could not parse JSON from LLM output (first 400 chars): %r", text[:400])
             raise ValueError(f"Could not parse JSON from VLM output: {text[:200]!r}")
         try:
-            data = json.loads(match.group(0))
+            data = json.loads(candidate)
         except json.JSONDecodeError as exc:
+            logger.warning("Brace-extracted JSON still invalid: %r", candidate[:400])
             raise ValueError(f"Could not parse JSON from VLM output: {exc}") from exc
 
     if isinstance(data, dict) and data.get("error") == "not_a_recipe":

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "kiwi"
-version = "0.6.0"
+version = "0.10.0"
 description = "Pantry tracking + leftover recipe suggestions"
 readme = "README.md"
 requires-python = ">=3.11"

scripts/pipeline/purple_carrot/discover_wayback.py

@@ -0,0 +1,298 @@
+"""
+discover_wayback.py — enumerate Purple Carrot recipe slugs via the Wayback Machine.
+
+Strategy:
+  1. CDX API  → all archived /api/v2/menus/* URLs (multiple timestamps)
+  2. Replay   → fetch each menu's menuItems, extract productPath slugs
+  3. CDX API  → all archived /api/v1/products/* URLs (direct slug capture)
+  4. CDX API  → /recipe-categories/* HTML pages for older slugs
+  5. Deduplicate and write manifest to OUT_FILE
+
+Output (JSONL, one record per recipe):
+  {"slug": "...", "title": "...", "subtitle": "...", "cook_time": "...",
+   "tags": [...], "serving_size": 2, "image_url": "...",
+   "wayback_ts": "20260412150557", "source": "menu|product_api|category_page"}
+
+Usage:
+  conda run -n cf python -m scripts.pipeline.purple_carrot.discover_wayback
+  conda run -n cf python -m scripts.pipeline.purple_carrot.discover_wayback --out /Library/Assets/kiwi/pipeline/pc_slugs.jsonl
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+import time
+from pathlib import Path
+from typing import Any
+from urllib.parse import urlencode
+
+import requests
+
+logger = logging.getLogger(__name__)
+
+CDX_BASE = "http://web.archive.org/cdx/search/cdx"
+WB_BASE = "https://web.archive.org/web"
+PC_HOST = "www.purplecarrot.com"
+
+# Polite delay between Wayback replay fetches (seconds)
+REPLAY_DELAY = 1.0
+CDX_DELAY = 0.5
+
+DEFAULT_OUT = Path("/Library/Assets/kiwi/pipeline/pc_slugs.jsonl")
+
+
+# ── CDX helpers ───────────────────────────────────────────────────────────────
+
+def cdx_query(url_pattern: str, **kwargs) -> list[dict]:
+    """Run a CDX search and return a list of result dicts."""
+    params = {
+        "url": url_pattern,
+        "output": "json",
+        "fl": "original,timestamp,statuscode",
+        "collapse": "urlkey",
+        "filter": "statuscode:200",
+        **kwargs,
+    }
+    for attempt in range(3):
+        try:
+            resp = requests.get(CDX_BASE, params=params, timeout=30)
+            resp.raise_for_status()
+            rows = resp.json()
+            if not rows or len(rows) < 2:
+                return []
+            headers = rows[0]
+            return [dict(zip(headers, row)) for row in rows[1:]]
+        except Exception as exc:
+            logger.warning("CDX attempt %d failed: %s", attempt + 1, exc)
+            time.sleep(2 ** attempt)
+    return []
+
+
+def wayback_get(url: str, timestamp: str) -> Any | None:
+    """Fetch a Wayback replay of a URL and return parsed JSON (or None)."""
+    replay_url = f"{WB_BASE}/{timestamp}/{url}"
+    for attempt in range(3):
+        try:
+            resp = requests.get(replay_url, timeout=30)
+            if resp.status_code == 200:
+                return resp.json()
+            if resp.status_code == 404:
+                return None
+        except Exception as exc:
+            logger.warning("Wayback GET attempt %d failed for %s: %s", attempt + 1, url, exc)
+        time.sleep(2 ** attempt)
+    return None
+
+
+# ── Slug extraction ───────────────────────────────────────────────────────────
+
+def slug_from_product_path(path: str) -> str | None:
+    """'/recipe/foo-bar-baz' → 'foo-bar-baz'."""
+    if not path:
+        return None
+    return path.strip("/").split("/")[-1] or None
+
+
+def _menu_item_to_record(item: dict, wayback_ts: str) -> dict | None:
+    slug = slug_from_product_path(item.get("productPath", ""))
+    if not slug:
+        return None
+    return {
+        "slug": slug,
+        "title": item.get("title", ""),
+        "subtitle": item.get("subtitle", ""),
+        "cook_time": item.get("cookTime", ""),
+        "tags": item.get("filterTags") or [],
+        "serving_size": item.get("servingSize"),
+        "image_url": item.get("imageURL", ""),
+        "description": item.get("description", ""),
+        "wayback_ts": wayback_ts,
+        "source": "menu",
+    }
+
+
+# ── Discovery passes ──────────────────────────────────────────────────────────
+
+def pass_menus(seen_slugs: set[str]) -> list[dict]:
+    """Walk all archived /api/v2/menus/* captures to extract slugs."""
+    records: list[dict] = []
+
+    # Find all distinct archived menu URLs
+    menu_cdx = cdx_query(f"{PC_HOST}/api/v2/menus/*", limit="500")
+    logger.info("CDX: %d archived menu URLs found", len(menu_cdx))
+    time.sleep(CDX_DELAY)
+
+    processed_menu_ids: set[str] = set()
+
+    for entry in menu_cdx:
+        url = entry["original"]
+        ts = entry["timestamp"]
+
+        # Skip the listing endpoint, only process individual menus
+        if not url.split("?")[0].rstrip("/").split("/")[-1].isdigit():
+            continue
+
+        menu_id = url.split("?")[0].rstrip("/").split("/")[-1]
+        if menu_id in processed_menu_ids:
+            continue
+        processed_menu_ids.add(menu_id)
+
+        logger.info("Fetching menu %s (ts=%s) ...", menu_id, ts)
+        data = wayback_get(url.split("?")[0] + "?logged_out=true", ts)
+        time.sleep(REPLAY_DELAY)
+
+        if not data or "menuItems" not in data:
+            continue
+
+        for item in data["menuItems"]:
+            rec = _menu_item_to_record(item, ts)
+            if rec and rec["slug"] not in seen_slugs:
+                seen_slugs.add(rec["slug"])
+                records.append(rec)
+                logger.debug("  + %s", rec["slug"])
+
+        logger.info("  %d new slugs (total so far: %d)", len(records), len(seen_slugs))
+
+    return records
+
+
+def pass_product_api(seen_slugs: set[str]) -> list[dict]:
+    """Pick up any directly archived /api/v1/products/* URLs the menu pass missed."""
+    records: list[dict] = []
+
+    product_cdx = cdx_query(f"{PC_HOST}/api/v1/products/*", limit="5000")
+    logger.info("CDX: %d archived product API URLs found", len(product_cdx))
+    time.sleep(CDX_DELAY)
+
+    for entry in product_cdx:
+        slug = entry["original"].rstrip("/").split("/")[-1]
+        if not slug or slug in seen_slugs:
+            continue
+        seen_slugs.add(slug)
+        records.append({
+            "slug": slug,
+            "title": "",
+            "subtitle": "",
+            "cook_time": "",
+            "tags": [],
+            "serving_size": None,
+            "image_url": "",
+            "description": "",
+            "wayback_ts": entry["timestamp"],
+            "source": "product_api",
+        })
+
+    logger.info("product_api pass: %d new slugs", len(records))
+    return records
+
+
+def pass_category_pages(seen_slugs: set[str]) -> list[dict]:
+    """Parse archived recipe-categories HTML pages for slugs not in the API.
+
+    Category pages are rendered SSR/with inline JSON state on older captures,
+    so we do a simple regex scan for /recipe/<slug> patterns.
+    """
+    import re
+
+    records: list[dict] = []
+    SLUG_RE = re.compile(r'["\s]/recipe/([a-z0-9][a-z0-9\-]{3,})["\s/?]')
+
+    cat_cdx = cdx_query(f"{PC_HOST}/recipe-categories/*", limit="200")
+    logger.info("CDX: %d archived category pages found", len(cat_cdx))
+    time.sleep(CDX_DELAY)
+
+    seen_category_urls: set[str] = set()
+
+    for entry in cat_cdx:
+        url = entry["original"].split("?")[0]
+        if url in seen_category_urls:
+            continue
+        seen_category_urls.add(url)
+
+        replay_url = f"{WB_BASE}/{entry['timestamp']}/{url}"
+        try:
+            resp = requests.get(replay_url, timeout=30)
+            time.sleep(REPLAY_DELAY)
+            if resp.status_code != 200:
+                continue
+        except Exception as exc:
+            logger.warning("Category page fetch failed: %s", exc)
+            continue
+
+        for slug in SLUG_RE.findall(resp.text):
+            if slug in seen_slugs:
+                continue
+            seen_slugs.add(slug)
+            records.append({
+                "slug": slug,
+                "title": "",
+                "subtitle": "",
+                "cook_time": "",
+                "tags": [],
+                "serving_size": None,
+                "image_url": "",
+                "description": "",
+                "wayback_ts": entry["timestamp"],
+                "source": "category_page",
+            })
+
+    logger.info("category_pages pass: %d new slugs", len(records))
+    return records
+
+
+# ── Main ──────────────────────────────────────────────────────────────────────
+
+def discover(out_file: Path) -> None:
+    seen: set[str] = set()
+
+    # Load previously discovered slugs so reruns are incremental
+    existing: list[dict] = []
+    if out_file.exists():
+        with open(out_file) as f:
+            for line in f:
+                line = line.strip()
+                if line:
+                    rec = json.loads(line)
+                    seen.add(rec["slug"])
+                    existing.append(rec)
+        logger.info("Loaded %d existing slugs from %s", len(seen), out_file)
+
+    new_records: list[dict] = []
+    new_records += pass_menus(seen)
+    new_records += pass_product_api(seen)
+    new_records += pass_category_pages(seen)
+
+    out_file.parent.mkdir(parents=True, exist_ok=True)
+    with open(out_file, "a") as f:
+        for rec in new_records:
+            f.write(json.dumps(rec) + "\n")
+
+    total = len(existing) + len(new_records)
+    logger.info(
+        "Done. %d new slugs written to %s (%d total).",
+        len(new_records), out_file, total,
+    )
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Discover Purple Carrot recipe slugs via Wayback")
+    parser.add_argument(
+        "--out",
+        type=Path,
+        default=DEFAULT_OUT,
+        help=f"Output JSONL manifest (default: {DEFAULT_OUT})",
+    )
+    parser.add_argument("--debug", action="store_true")
+    args = parser.parse_args()
+
+    logging.basicConfig(
+        level=logging.DEBUG if args.debug else logging.INFO,
+        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+    )
+
+    discover(args.out)
+
+
+if __name__ == "__main__":
+    main()

scripts/pipeline/purple_carrot/scrape_recipes.py

@@ -0,0 +1,529 @@
+"""
+scrape_recipes.py — fetch full recipe data for slugs in pc_slugs.jsonl.
+
+For each slug:
+  1. Try Wayback /api/v1/products/<slug> — oldest capture first (pre-HelloFresh
+     acquisition data is more complete).
+  2. If instructions are empty, try the recipe HTML page via Wayback and parse
+     inline JSON state or structured markup.
+  3. Merge with metadata already in the manifest (title, tags, cook_time, etc.)
+  4. Emit one row per recipe to recipes_purplecarrot.parquet in food.com columnar
+     format so build_recipe_index.py can import it unchanged.
+
+Output columns (food.com schema + PC extras ignored by the indexer):
+  RecipeId, Name, Subtitle, RecipeIngredientParts, RecipeInstructions,
+  RecipeCategory, Keywords, Calories, FatContent, ProteinContent,
+  SodiumContent, SugarContent, CarbohydrateContent, FiberContent,
+  RecipeServings, Description, ImageURL, CookTime, Slug, Source
+
+Usage:
+  conda run -n cf python -m scripts.pipeline.purple_carrot.scrape_recipes
+  conda run -n cf python -m scripts.pipeline.purple_carrot.scrape_recipes \\
+      --slugs /Library/Assets/kiwi/pipeline/pc_slugs.jsonl \\
+      --out   /Library/Assets/kiwi/pipeline/recipes_purplecarrot.parquet \\
+      --resume
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+import re
+import time
+from pathlib import Path
+from typing import Any
+
+import requests
+
+logger = logging.getLogger(__name__)
+
+CDX_BASE = "http://web.archive.org/cdx/search/cdx"
+WB_BASE = "https://web.archive.org/web"
+PC_HOST = "www.purplecarrot.com"
+
+REPLAY_DELAY = 1.2
+CDX_DELAY = 0.5
+
+DEFAULT_SLUGS = Path("/Library/Assets/kiwi/pipeline/pc_slugs.jsonl")
+DEFAULT_OUT = Path("/Library/Assets/kiwi/pipeline/recipes_purplecarrot.parquet")
+
+# Inline JSON state embedded by the SSR renderer — used as fallback HTML parser
+_NEXT_DATA_RE = re.compile(r'<script id="__NEXT_DATA__"[^>]*>(.*?)</script>', re.DOTALL)
+_REDUX_STATE_RE = re.compile(r'window\.__INITIAL_STATE__\s*=\s*(\{.*?\});\s*\n', re.DOTALL)
+
+
+# ── Wayback helpers ───────────────────────────────────────────────────────────
+
+def _cdx_timestamps(slug: str) -> list[str]:
+    """Return all captured timestamps for a product slug, oldest first."""
+    url = f"{PC_HOST}/api/v1/products/{slug}"
+    try:
+        resp = requests.get(
+            CDX_BASE,
+            params={
+                "url": url,
+                "output": "json",
+                "fl": "timestamp,statuscode",
+                "filter": "statuscode:200",
+                "limit": "20",
+            },
+            timeout=20,
+        )
+        resp.raise_for_status()
+        rows = resp.json()
+        if len(rows) < 2:
+            return []
+        return [row[0] for row in rows[1:]]  # timestamps only, oldest first
+    except Exception as exc:
+        logger.debug("CDX timestamps failed for %s: %s", slug, exc)
+        return []
+
+
+def _wayback_json(url: str, timestamp: str) -> Any | None:
+    replay = f"{WB_BASE}/{timestamp}/{url}"
+    for attempt in range(3):
+        try:
+            resp = requests.get(replay, timeout=30)
+            if resp.status_code == 200:
+                return resp.json()
+            if resp.status_code in (404, 410):
+                return None
+        except Exception as exc:
+            logger.debug("Wayback JSON attempt %d failed (%s): %s", attempt + 1, url, exc)
+        time.sleep(2 ** attempt)
+    return None
+
+
+def _wayback_html(url: str, timestamp: str) -> str | None:
+    replay = f"{WB_BASE}/{timestamp}/{url}"
+    for attempt in range(3):
+        try:
+            resp = requests.get(replay, timeout=30)
+            if resp.status_code == 200:
+                return resp.text
+            if resp.status_code in (404, 410):
+                return None
+        except Exception as exc:
+            logger.debug("Wayback HTML attempt %d failed (%s): %s", attempt + 1, url, exc)
+        time.sleep(2 ** attempt)
+    return None
+
+
+# ── Recipe extraction from API JSON ──────────────────────────────────────────
+
+def _extract_from_api(data: dict) -> dict | None:
+    """Parse a /api/v1/products/<slug> response into our recipe dict.
+
+    Returns None if the response has no usable content (empty title, etc.).
+    Returns a partial dict if only some fields are populated — caller merges
+    with manifest metadata.
+    """
+    if not data or not isinstance(data, dict):
+        return None
+
+    title = data.get("title", "").strip()
+    subtitle = data.get("subtitle", "").strip()
+    slug = data.get("slug", "")
+
+    skus = data.get("skus") or []
+    sku = skus[0] if skus else {}
+
+    # Instructions: list of {step_number, title, description}
+    raw_instructions = sku.get("instructions") or []
+    steps: list[str] = []
+    for step in sorted(raw_instructions, key=lambda s: s.get("step_number", 0)):
+        parts = []
+        if step.get("title"):
+            parts.append(step["title"])
+        if step.get("description"):
+            parts.append(step["description"])
+        if parts:
+            steps.append(". ".join(parts))
+
+    # Ingredients: may be in ingredients_quantity or ingredients
+    raw_ingr = sku.get("ingredients_quantity") or sku.get("ingredients") or []
+    ingredients: list[str] = []
+    for item in raw_ingr:
+        if isinstance(item, dict):
+            qty = item.get("quantity") or item.get("qty") or ""
+            unit = item.get("unit") or ""
+            name = item.get("name") or item.get("ingredient", {}).get("name", "") if isinstance(item.get("ingredient"), dict) else item.get("ingredient_name", "")
+            raw = item.get("raw") or item.get("display_name") or ""
+            line = raw or " ".join(filter(None, [str(qty), str(unit), str(name)])).strip()
+            if line:
+                ingredients.append(line)
+        elif isinstance(item, str) and item.strip():
+            ingredients.append(item.strip())
+
+    nutrition = sku.get("nutrition_label") or {}
+    calories = _num(nutrition.get("calories") or sku.get("calories"))
+    fat = _num(nutrition.get("total_fat") or sku.get("fat"))
+    protein = _num(nutrition.get("protein") or sku.get("protein"))
+    sodium = _num(nutrition.get("sodium") or sku.get("sodium"))
+    sugar = _num(nutrition.get("sugar") or nutrition.get("total_sugars"))
+    carbs = _num(nutrition.get("total_carbohydrate") or sku.get("carbs"))
+    fiber = _num(nutrition.get("dietary_fiber") or sku.get("fiber"))
+
+    tags = sku.get("tags") or data.get("tags") or []
+    category = sku.get("meal_type") or sku.get("product_type") or ""
+    servings = _num(sku.get("servings"))
+
+    cook_time = sku.get("prep_and_cook_time") or ""
+    description = sku.get("description") or ""
+
+    images = sku.get("hero_images") or sku.get("image_versions") or []
+    image_url = ""
+    if images and isinstance(images[0], dict):
+        image_url = images[0].get("image_url") or images[0].get("url") or ""
+    if not image_url and data.get("square_image"):
+        sq = data["square_image"]
+        image_url = sq.get("url") if isinstance(sq, dict) else ""
+
+    return {
+        "slug": slug,
+        "title": title,
+        "subtitle": subtitle,
+        "steps": steps,
+        "ingredients": ingredients,
+        "category": category,
+        "tags": tags,
+        "calories": calories,
+        "fat": fat,
+        "protein": protein,
+        "sodium": sodium,
+        "sugar": sugar,
+        "carbs": carbs,
+        "fiber": fiber,
+        "servings": servings,
+        "cook_time": cook_time,
+        "description": description,
+        "image_url": image_url,
+        "has_full_recipe": bool(steps and ingredients),
+    }
+
+
+def _num(val: Any) -> float | None:
+    if val is None:
+        return None
+    try:
+        v = float(str(val).replace("g", "").replace("mg", "").split()[0])
+        return v if v > 0 else None
+    except Exception:
+        return None
+
+
+# ── Fallback: HTML inline state parsing ──────────────────────────────────────
+
+def _extract_from_html(html: str, slug: str) -> dict | None:
+    """Try to pull recipe data from inline JS state in older SSR pages."""
+    # Attempt 1: Next.js __NEXT_DATA__
+    m = _NEXT_DATA_RE.search(html)
+    if m:
+        try:
+            state = json.loads(m.group(1))
+            # Walk the Next.js page props tree looking for recipe data
+            props = state.get("props", {}).get("pageProps", {})
+            recipe = props.get("recipe") or props.get("product")
+            if recipe and isinstance(recipe, dict) and recipe.get("title"):
+                return _extract_from_api(recipe)
+        except Exception:
+            pass
+
+    # Attempt 2: Redux __INITIAL_STATE__
+    m = _REDUX_STATE_RE.search(html)
+    if m:
+        try:
+            state = json.loads(m.group(1))
+            # Try common Redux state shapes
+            for key in ("recipe", "product", "currentRecipe", "currentProduct"):
+                recipe = state.get(key)
+                if recipe and isinstance(recipe, dict) and recipe.get("title"):
+                    return _extract_from_api(recipe)
+        except Exception:
+            pass
+
+    # Attempt 3: JSON-LD structured data
+    ld_matches = re.findall(
+        r'<script[^>]+type=["\']application/ld\+json["\'][^>]*>(.*?)</script>',
+        html, re.DOTALL
+    )
+    for raw in ld_matches:
+        try:
+            ld = json.loads(raw)
+            if isinstance(ld, list):
+                ld = next((x for x in ld if x.get("@type") == "Recipe"), None)
+            if not ld or ld.get("@type") != "Recipe":
+                continue
+            steps = []
+            for inst in (ld.get("recipeInstructions") or []):
+                if isinstance(inst, dict):
+                    steps.append(inst.get("text", ""))
+                elif isinstance(inst, str):
+                    steps.append(inst)
+            ingredients = ld.get("recipeIngredient") or []
+            return {
+                "slug": slug,
+                "title": ld.get("name", ""),
+                "subtitle": "",
+                "steps": [s for s in steps if s],
+                "ingredients": [i for i in ingredients if i],
+                "category": ld.get("recipeCategory", ""),
+                "tags": ld.get("keywords", "").split(",") if isinstance(ld.get("keywords"), str) else [],
+                "calories": _num((ld.get("nutrition") or {}).get("calories")),
+                "fat": None, "protein": None, "sodium": None,
+                "sugar": None, "carbs": None, "fiber": None,
+                "servings": _num(ld.get("recipeYield")),
+                "cook_time": str(ld.get("totalTime") or ld.get("cookTime") or ""),
+                "description": ld.get("description", ""),
+                "image_url": (ld["image"][0] if isinstance(ld.get("image"), list) else ld.get("image", "")) or "",
+                "has_full_recipe": True,
+            }
+        except Exception:
+            pass
+
+    return None
+
+
+# ── Per-slug fetch ─────────────────────────────────────────────────────────────
+
+def fetch_recipe(slug: str, manifest_meta: dict) -> dict | None:
+    """Fetch the fullest available recipe data for a slug from Wayback.
+
+    Returns a merged dict of manifest metadata + API/HTML-extracted content.
+    """
+    api_url = f"https://{PC_HOST}/api/v1/products/{slug}"
+    html_url = f"https://{PC_HOST}/recipe/{slug}"
+
+    recipe: dict | None = None
+
+    # Try product API — oldest captures are most likely to have full data
+    timestamps = _cdx_timestamps(slug)
+    time.sleep(CDX_DELAY)
+
+    if not timestamps and manifest_meta.get("wayback_ts"):
+        timestamps = [manifest_meta["wayback_ts"]]
+
+    for ts in timestamps:
+        data = _wayback_json(api_url, ts)
+        time.sleep(REPLAY_DELAY)
+        if not data:
+            continue
+        candidate = _extract_from_api(data)
+        if not candidate:
+            continue
+        recipe = candidate
+        if recipe.get("has_full_recipe"):
+            logger.debug("[%s] Full recipe from API (ts=%s)", slug, ts)
+            break
+        logger.debug("[%s] Partial API data (ts=%s) — trying HTML fallback", slug, ts)
+
+    # HTML fallback when API has no steps/ingredients
+    if not recipe or not recipe.get("has_full_recipe"):
+        html_cdx_url = f"{PC_HOST}/recipe/{slug}"
+        try:
+            html_resp = requests.get(
+                CDX_BASE,
+                params={
+                    "url": html_cdx_url,
+                    "output": "json",
+                    "fl": "timestamp,statuscode",
+                    "filter": "statuscode:200",
+                    "limit": "5",
+                },
+                timeout=20,
+            )
+            html_ts_rows = html_resp.json() if html_resp.ok else []
+            html_timestamps = [row[0] for row in html_ts_rows[1:]] if len(html_ts_rows) > 1 else []
+        except Exception:
+            html_timestamps = []
+        time.sleep(CDX_DELAY)
+
+        for ts in html_timestamps:
+            html = _wayback_html(html_url, ts)
+            time.sleep(REPLAY_DELAY)
+            if not html:
+                continue
+            html_recipe = _extract_from_html(html, slug)
+            if html_recipe and html_recipe.get("has_full_recipe"):
+                logger.debug("[%s] Full recipe from HTML (ts=%s)", slug, ts)
+                recipe = html_recipe
+                break
+
+    # Build merged record: manifest metadata fills any gaps from API/HTML
+    merged: dict = {
+        "slug": slug,
+        "title": manifest_meta.get("title", ""),
+        "subtitle": manifest_meta.get("subtitle", ""),
+        "steps": [],
+        "ingredients": [],
+        "category": "",
+        "tags": manifest_meta.get("tags") or [],
+        "calories": None,
+        "fat": None,
+        "protein": None,
+        "sodium": None,
+        "sugar": None,
+        "carbs": None,
+        "fiber": None,
+        "servings": manifest_meta.get("serving_size"),
+        "cook_time": manifest_meta.get("cook_time", ""),
+        "description": manifest_meta.get("description", ""),
+        "image_url": manifest_meta.get("image_url", ""),
+        "source": "purple_carrot",
+        "wayback_ts": manifest_meta.get("wayback_ts", ""),
+        "has_full_recipe": False,
+    }
+
+    if recipe:
+        for key in recipe:
+            # Prefer API/HTML data; keep manifest value only when API field is empty
+            val = recipe[key]
+            if val or key not in merged or not merged[key]:
+                merged[key] = val
+
+    if not merged["title"]:
+        logger.warning("[%s] No title — skipping", slug)
+        return None
+
+    return merged
+
+
+# ── Output formatting ─────────────────────────────────────────────────────────
+
+def _to_dataframe_row(r: dict) -> dict:
+    """Convert merged recipe dict to food.com-compatible parquet row."""
+    # Build plain-text input for allrecipes-style corpus compatibility
+    lines = [r["title"]]
+    if r.get("subtitle"):
+        lines.append(r["subtitle"])
+    if r.get("description"):
+        lines.append("")
+        lines.append(r["description"])
+    if r.get("ingredients"):
+        lines += ["", "Ingredients:"] + [f"- {i}" for i in r["ingredients"]]
+    if r.get("steps"):
+        lines += ["", "Directions:"] + [f"- {s}" for s in r["steps"]]
+    plain_text = "\n".join(lines)
+
+    source_url = f"https://www.purplecarrot.com/recipe/{r['slug']}"
+
+    return {
+        # food.com schema columns (used by build_recipe_index.py)
+        "RecipeId": f"pc_{r['slug']}",
+        "Name": r["title"],
+        "RecipeIngredientParts": r.get("ingredients") or [],
+        "RecipeInstructions": r.get("steps") or [],
+        "RecipeCategory": r.get("category", ""),
+        "Keywords": r.get("tags") or [],
+        "Calories": r.get("calories"),
+        "FatContent": r.get("fat"),
+        "ProteinContent": r.get("protein"),
+        "SodiumContent": r.get("sodium"),
+        "SugarContent": r.get("sugar"),
+        "CarbohydrateContent": r.get("carbs"),
+        "FiberContent": r.get("fiber"),
+        "RecipeServings": r.get("servings"),
+        # PC-specific extras (ignored by indexer, used by training pipeline)
+        "Subtitle": r.get("subtitle", ""),
+        "Description": r.get("description", ""),
+        "ImageURL": r.get("image_url", ""),
+        "CookTime": r.get("cook_time", ""),
+        "Slug": r["slug"],
+        "Source": "purple_carrot",
+        "SourceURL": source_url,       # canonical attribution link shown in recipe UI
+        "HasFullRecipe": r.get("has_full_recipe", False),
+        "WaybackTs": r.get("wayback_ts", ""),
+        # Also emit plain-text input for allrecipes-compatible corpus search
+        "input": plain_text,
+    }
+
+
+# ── Main ──────────────────────────────────────────────────────────────────────
+
+def scrape(slugs_file: Path, out_file: Path, resume: bool = True) -> None:
+    import pandas as pd
+
+    # Load manifest
+    if not slugs_file.exists():
+        logger.error("Slugs manifest not found: %s", slugs_file)
+        return
+
+    manifest: dict[str, dict] = {}
+    with open(slugs_file) as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                rec = json.loads(line)
+                slug = rec["slug"]
+                # Keep the richest metadata if slug appears from multiple sources
+                if slug not in manifest or rec.get("source") == "menu":
+                    manifest[slug] = rec
+
+    logger.info("Manifest: %d unique slugs", len(manifest))
+
+    # Load already-scraped slugs for resume
+    done_slugs: set[str] = set()
+    existing_rows: list[dict] = []
+    if resume and out_file.exists():
+        try:
+            existing_df = pd.read_parquet(out_file)
+            done_slugs = set(existing_df["Slug"].tolist())
+            existing_rows = existing_df.to_dict("records")
+            logger.info("Resume: %d already scraped", len(done_slugs))
+        except Exception as exc:
+            logger.warning("Could not load existing parquet for resume: %s", exc)
+
+    todo = [s for s in manifest if s not in done_slugs]
+    logger.info("%d slugs to fetch", len(todo))
+
+    rows = list(existing_rows)
+    for i, slug in enumerate(todo, 1):
+        logger.info("[%d/%d] %s", i, len(todo), slug)
+        recipe = fetch_recipe(slug, manifest[slug])
+        if recipe:
+            rows.append(_to_dataframe_row(recipe))
+            status = "full" if recipe.get("has_full_recipe") else "partial"
+            logger.info("  -> %s (%s)", recipe.get("title", "?"), status)
+        else:
+            logger.warning("  -> skipped (no title)")
+
+        # Write checkpoint every 50 recipes
+        if i % 50 == 0:
+            _write_parquet(rows, out_file)
+            logger.info("Checkpoint: %d recipes written", len(rows))
+
+    _write_parquet(rows, out_file)
+    full = sum(1 for r in rows if r.get("HasFullRecipe"))
+    logger.info(
+        "Done. %d recipes written to %s (%d full, %d partial).",
+        len(rows), out_file, full, len(rows) - full,
+    )
+
+
+def _write_parquet(rows: list[dict], out_file: Path) -> None:
+    import pandas as pd
+    out_file.parent.mkdir(parents=True, exist_ok=True)
+    pd.DataFrame(rows).to_parquet(out_file, index=False)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Scrape Purple Carrot recipes from Wayback")
+    parser.add_argument("--slugs", type=Path, default=DEFAULT_SLUGS)
+    parser.add_argument("--out", type=Path, default=DEFAULT_OUT)
+    parser.add_argument(
+        "--no-resume", dest="resume", action="store_false",
+        help="Start fresh (ignore existing parquet)",
+    )
+    parser.add_argument("--debug", action="store_true")
+    args = parser.parse_args()
+
+    logging.basicConfig(
+        level=logging.DEBUG if args.debug else logging.INFO,
+        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+    )
+
+    scrape(args.slugs, args.out, resume=args.resume)
+
+
+if __name__ == "__main__":
+    main()