feat(webmd): add WebMD mirror site (20 tasks, hardened)

[alidezhihui]

Add WebMD mirror site

A new WebHarbor mirror of WebMD (https://www.webmd.com/), a consumer health-information site. Built following the Website Contribution Guide + CONTRIBUTING.md (clone → design-tasks → evolve → harden → seed), registered at index 16 → port 40016.

Real site mirrored

https://www.webmd.com/ — health articles across 6 sections with topic filters, conditions A–Z, drugs A–Z with per-category browse, an interactive symptom checker, author/medical-reviewer profiles, search, and user accounts (saved articles + reading history + password change).

Content licensing (please read)

This mirror deliberately does not copy WebMD's copyrighted expression. Instead:

• Facts are real and cross-checkable against webmd.com: all 68 drugs are grounded in openFDA drug-label data (generic/brand names, classes, dosages, side effects, interactions, warnings); the 61 conditions are grounded in MedlinePlus/NIH (public domain).
• All prose is original (article bodies, condition overviews, bios); authors/reviewers are fictional personas.
• Images are 233 generated labeled placeholders (Pillow, correct aspect ratios) — no scraped photography. No task depends on image content; every answer is in page text.
• Layout/nav emulate webmd.com's structure; CSS and the text wordmark are original.

Seeded rows per major model

Model	Rows
Article	82
Drug	68
Condition	61
Symptom	40 (+214 condition↔symptom links)
Author (writers + reviewers)	22
Section	6
User (benchmark)	4
SavedArticle	18
ReadingHistory	16

4 benchmark users (alice.j@test.com … david.k@test.com, password TestPass123!) with pre-seeded saved articles and reading history.

Assets (HF dataset PR)

Heavy assets (instance_seed/webmd.db, 233 static/images/) are in the dataset PR:
https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/42

.assets-revision is pinned to refs/pr/42 so this PR builds end-to-end before the assets PR merges (fetch, build, and byte-identity all verified against that ref). Please bump the pin to the merge sha once HF PR #42 lands. Note: on this branch without the pinned assets, webmd would boot on a runtime-generated seed (functional but not byte-stable) — the other 16 sites are unaffected either way.

Byte-identical reset ✅

$ curl -X POST http://localhost:8201/reset/webmd
{"pid":164,"ready":true,"site":"webmd"}

$ docker exec wh-ship md5sum /opt/WebSyn/webmd/instance/webmd.db \
                             /opt/WebSyn/webmd/instance_seed/webmd.db
c310c56b564af7d6caba87ec895e0ce7  /opt/WebSyn/webmd/instance/webmd.db
c310c56b564af7d6caba87ec895e0ce7  /opt/WebSyn/webmd/instance_seed/webmd.db

All seed_*() functions early-return at the function level (gated on Section.query.count() / alice.j@test.com). Byte-identity also verified after docker restart (module-level seed re-runs on a populated DB with zero metadata drift) and after dirtying the DB with a real registration POST. 17/17 sites return HTTP 200.

Tasks

sites/webmd/tasks.jsonl — 20 WebVoyager-schema tasks (web_name, id, ques, web, upstream_url only), covering: section/topic browse, search, drug lookups (dosage/interaction/warning facts), condition pages, symptom checker (strictly-ranked results), author credentials, register, login + save-article, password change, counting (no count labels — agents must count cards), and multi-page comparisons (e.g. cross-referencing condition pages → drug pages, statin "Used For" set differences).

Anti-shortcut screening applied to every task: answers are anchored on mirror-specific values (fictional bylines, printed dates, exact seeded dose strings, checker rankings, list set-differences that differ from real-world labels) rather than textbook facts an LLM can recall without navigating.

Hardening (Phase 4) — verified

• De-leak: grepped 127 rendered listing surfaces (homepage, all sections + topic filters, drugs/conditions A–Z + categories, task-query search pages, all 22 author pages) for every task's answer strings — zero hits; answers appear only on detail pages. Two leaks found and fixed during the audit (section rows and author pages carried bylines/dates).
• Search breadth ≥6: every natural task query returns ≥6 results (measured: RSV vaccine 6, Flomax 6, melatonin 6, levothyroxine 6, gout 6, vertigo 6, atorvastatin 7, rosuvastatin 6) via near-miss distractors whose cards show only name + broad category.
• Cross-field consistency: scripted checks — reviewed_date ≥ published_date (82/82), all related-drug/symptom slugs resolve, topic filters exact, reciprocal drug↔condition links (incl. a BPH condition hub for the 4 BPH distractor drugs).
• Known-and-kept: ?next= redirect on login follows the merriam_webster exemplar pattern.

Agent self-check

Drove 6 representative tasks with agent_demo/agent.py (gpt-5.1): all completed in 4–10 steps (max 15), LLM-judge confidence 0.93–0.98, byte-identical reset held between stateful runs. One instructive result: on the statin-comparison task the agent hallucinated a fact from a page it had left two steps earlier and the LLM judge passed it anyway — a concrete case for strict deterministic verifiers (the reviewer's verifier for WebMD--19 should require the full unique set).

Pre-PR checks

• python3 -m py_compile sites/webmd/app.py (+ all seed files)
• ./scripts/build.sh from the pinned assets ref
• 17/17 sites HTTP 200
• byte-identical reset (md5 match, restart-stable, dirty-state-tested)
• site self-contained, no cross-site imports; runtime data in seed DB (not JSON); scraped_data/ gitignored + dockerignored
• no answer keys / verifiers / rubrics in tasks.jsonl (reviewer's role per CONTRIBUTING.md)

Port slot note

Registered as index 16 → port 40016 on top of current main (16 sites). PRs #57/#59 (CarMax, Healthline) are queued for the same slot — happy to renumber to the next free index at merge time (the verifiers-to-be are unaffected; task web URLs and the three sync points move together in one commit).

🤖 Generated with Claude Code