Add TED mirror site

[shanjiaming]

Summary

• Adds a TED mirror site with browse, search, topic, playlist, event, auth, account, and saved-talk workflows.
• Seeds 64 real TED talks with real thumbnails, 7 playlists, 17 event contexts, and 4 benchmark users.
• Registers TED as the 17th WebHarbor site on port 40016.
• Addresses review feedback by hiding answer fields on listing cards, tightening search matching, adding near-miss distractors, correcting TEDx cities, and removing unrelated PR noise.

Assets

• HF asset tarball: ted.tar.gz
• HF dataset revision: https://huggingface.co/datasets/ChilleD/WebHarbor/commit/597623a2f32898afa12e3bbeda15520f559aa7c7
• HF asset PR: https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/2 (open; maintainer merge is still required before final approval).
• Contains sites/ted/instance_seed/ted.db and sites/ted/static/images/.
• Seed DB MD5: 5fcca409b441876f22b8964ff681be91

Verification

• python3 -m py_compile sites/ted/app.py sites/ted/seed_data.py control_server.py site_runner.py
• ./scripts/check_assets.sh
• git diff --check
• ./scripts/fetch_assets.sh ted
• ./scripts/build.sh webharbor:ted-final-review
• Docker run with -p 8401:8101 -p 43000-43016:40000-40016.
• All ports 43000-43016 returned HTTP 200.
• TED _health returned {ok: true, site: ted, talks: 64}.
• TED runtime audit covered listing answer leaks, detail answers, login, profile update, saving talks, event registration, new account flow, /reset/ted, byte-identical DB restore, and /reset-all.
• Browser audit covered desktop and 390px mobile layouts across home, search, topic, playlist, detail, and events pages with no broken images, no console warnings/errors, and no horizontal overflow.

[Susan571]

Reviewer checklist (WebHarbor Review Guide)

Reviewer note: Source-level review (checked out pull/1/head). The mechanical Docker pipeline was not run by me — claims (all 16 ports 200, seed md5 8ba9c3da… byte-identical, /reset-all ~1.1s) are plausible; please re-verify in a Docker-capable pass before merge.

Visual fidelity: ✅ PASS (with note)

• Uses real TED data: 64 real talks with real titles, speakers, events (TED2026/TEDNext 2025/TEDx…), durations, view counts, topics, canonical ted.com URLs, and real compressed thumbnails shipped via the paired HF asset (static/images/talk_*.jpg). Templates render those real images on cards and detail pages — a real mirror, not a placeholder skin.
• Note: transcript fields are templated boilerplate ("opens by framing the central tension…"), not real transcripts — acceptable, but don't write tasks that depend on transcript content.

Functional depth: ✅ PASS

• Browse/filter talks (event + max-minutes), topics, playlists, events, search, talk detail (duration/views/topics/related), register, login, save talk with note, remove saved talk, newsletter-topic update, event interest registration. Real auth + multiple CRUD/state flows.

Task quality: ⚠️ Weak pass (18 tasks — meets the count)

• Strong, varied, multi-step: TED--15 (two music talks, which has more views → views appear only on the detail page, so both must be opened, no leak), --8 (compare two durations), --10/--3 (playlist reasoning), --16 (create account → save → confirm), --12 (remove a non-AI saved talk).
• Mild leak: talk cards show event - {{minutes}} min, so duration-in-minutes tasks (--0, --2, --8, --11) are partly answerable from the listing. --0 instructs "open the talk … what is the duration," so navigation is forced, but consider hiding minutes on cards for duration tasks.
• Credentials: tasks reference alice.j@test.com but generally not the password — better than the other PRs.

Verdict: Comment — approve after minor fixes + mechanical re-verify

(1) Remove/relocate the minutes-on-card leak for duration tasks; (2) a Docker-capable reviewer re-runs build + 200s + byte-identical reset before final merge. This is the strongest of the five environments I reviewed.

[MufanQiu]

Review — ted (PR #1)

Verdict: REQUEST CHANGES.

We deployed the environment locally and ran the full review checklist; every finding below was independently reproduced, and fidelity was checked against the real upstream site.

Issues (summary)

• BLOCKER: .assets-revision pinned to 'main' but new site's tarball is not on main -> CI fail-closed
• MINOR: Event-name + title exposed on every talk card leaks the answer for TED--3, TED--6, TED--10
• MINOR: Thin single-result searches make TED--13 and TED--14 trivially easy
• MINOR: scored_talks over-matches short substrings; q='AI' returns all 64 talks
• MINOR: Ambiguous referents in a few task wordings (TED--8 two Debbie Millman talks; TED--2/15 'a/two' loosely specified)
• NIT: Out-of-scope repo edits bundled in a single-site PR (review-env SKILL.md, agent_demo, CONTRIBUTING, booking image paths)

Mechanical checks: FAIL

The site registration is internally consistent: EXPOSE 8101 40000-40015 (Dockerfile:36), the web field http://localhost:40015/ on all 18 task rows, and ted at index 15 in both websyn_start.sh SITES and control_server.py SITES all agree.

Visual fidelity: PASS

Strong TED.com resemblance across every page. Homepage: red TED wordmark, "Ideas change everything" hero with featured-talk thumbnail, WATCH/TOPICS/PLAYLISTS/ATTEND nav, "Search ideas" box, "Latest talks" card grid (speaker + thumbnail + topic chips), "Popular now" with view counts, "Curated playlists", topic-tag footer; title TED: Ideas change everything.

Functional depth: PASS

All flows verified via Playwright. Login (alice.j@test.com / TestPass123!) → /account (H1 "Alice Johnson"); invalid password rejected ("Invalid email or password"); register auto-logs-in, duplicate email rejected ("already exists").

Task quality: FAIL

18 tasks, all with confirmed DB ground truth; catalog is rich (64 talks, 17 events, 124 topics, 7 playlists) and difficulty distribution meets the bar (≥5 multi-action tasks, ≥2 comparisons).

Required fixes before approval

1. (BLOCKER) Re-pin .assets-revision from main back to the merged HuggingFace commit SHA that contains sites/ted/ assets, set after HF PR feat(mega): add new site #2 is merged.
2. (MAJOR) Eliminate the listing-card answer leaks for TED--3, TED--6, TED--10, and TED--5.
3. (MINOR) Add near-miss distractors so TED--13 (wine) and TED--14 (architecture 3D printing) are not single-result searches with the answer on the only card.
4. (MINOR) Fix scored_talks (app.py:155-165) to use word-boundary/tokenized matching or require token length ≥ 3 so q=AI no longer matches all 64 talks via the ai substring.
5. (MINOR) Correct the six TEDx event cities (currently all "Washington, DC") to their real locations (Manchester, Denver, Berlin, etc.) so /events data is faithful.