Align YouTube and Weather snapshot fidelity

[YurunChen]

Summary

• Aligns sites/youtube and sites/weather routes, templates, seeds, and tasks for snapshot-style fidelity and deterministic benchmark flows.
• Fixes .gitignore patterns for sites/*/instance/, sites/*/scraped_data/, and agent_demo/.playwright/.

HF assets

Heavy assets (instance_seed/*.db, static/images/, etc.) ship via the Hugging Face dataset per CONTRIBUTING.md. If this PR changes bundled assets, pair it with an HF dataset PR and bump .assets-revision to the merged dataset revision.

Verification

• python3 -m py_compile on touched app.py files
• Docker: control /health, key routes 200, POST /reset/<site> with matching md5sum for instance/*.db vs instance_seed/*.db for youtube and weather.

Made with Cursor

[MufanQiu]

Review — weather (PR #18)

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 pins a main SHA that almost certainly lacks weather.tar.gz (CI fail-closed)
• MAJOR: Answer leak: search results cards expose current temperature and condition label
• MAJOR: Search has no distractors: every city query returns exactly one result (click-result-Add TED mirror site #1)
• MAJOR: Task 11 is ill-posed: radar image is a generic shared US map, not city-specific
• MINOR: 514 AppleDouble ._* junk files + unused scraped/ image dir ship in the tarball
• MINOR: Committed seed_data.py cannot regenerate the shipped seed DB (code/data drift)

Mechanical checks: FAIL

Registration is internally consistent across all three sync points: weather is index 16 in websyn_start.sh SITES and control_server.py SITES, giving port 40000+16 = 40016; the Dockerfile EXPOSE 8101 40000-40016 and the tasks' web field (http://localhost:40016/) agree.

Visual fidelity: PASS

Strong fidelity to weather.com. Homepage shows the Weather Channel logo, grouped left-sidebar nav, the New York current-conditions card (41°F Cloudy, Day 43 / Night 35), Health & Activities modules, Top Stories hero + tiles, Video list, and right rail.

Functional depth: PASS

All flows verified with Playwright. Login works for the 4 seeded benchmark users (TestPass123!). Unit preference is correctly user-aware: carol.d (metric) sees Tokyo 14°C; bob.c (imperial) sees Tokyo 58°F (correct conversion).

Task quality: FAIL

20 tasks; every task's ground truth is reachable in the rendered UI, but the set fails the quality bar on three cross-cutting axes — detailed in the issue summary above and the required fixes below.

Required fixes before approval

1. (BLOCKER, F1) Re-pin wt/weather/.assets-revision to a HF main SHA that actually contains weather.tar.gz (i.e. main after dataset PR Add UC Berkeley mirror site (port 40015) #10 merges).
2. (BLOCKER, F2) Remove the leak in templates/search.html line 20.
3. (MAJOR, F3) Add near-miss / same-name distractor locations so city searches return ≥6 mixed results with the target not at position Add TED mirror site #1.
4. (MAJOR, F4) Fix Task 11.
5. (MAJOR, F5) Strip all 514 AppleDouble ._* files (including instance_seed/._weather.db) before packing the tarball (COPYFILE_DISABLE=1 tar / find -name '._*' -delete) and delete the unreferenced static/images/weather/scraped/ directory.