fix(amazon): improve storefront fidelity and task grading

[jackjin1997]

Summary

• replace the legacy generic hero with a horizontally browsable promotional deck
• bring the navigation, homepage cards, spacing, and footer closer to the current Amazon storefront structure
• curate the legacy 41-task set to 20 executable tasks with explicit outputs, aligned rubrics, and one deterministic verifier per task
• enforce seed-derived facts, target-origin interaction evidence, read-only database invariants, and exact wishlist/cart state deltas
• add one pinned Fire TV hero asset through the paired Hugging Face PR #81

Visual reference

Compared against https://www.amazon.com/ on 2026-09-13. Amazon campaign content and localization rotate dynamically, so this change targets stable structure and visual hierarchy rather than claiming pixel-identical promotional content.

Task contract

• accepted 20 tasks after browser execution and seed validation: 18 read-only and 2 authenticated state tasks
• preserved original task IDs; weak, redundant, randomized, or time-relative tasks were removed rather than renumbered or graded loosely
• wishlist and cart verifiers compare initial/after SQLite snapshots and reject wrong accounts, wrong products, no-ops, and unrelated mutations
• list/ranking verifiers bind values to products and reject swapped values, missing entries, and extra invalid products
• qualifying alternative products are accepted where the prompt permits them
• search evidence compares numeric filter values numerically, so equivalent values such as 25, 25.0, and 25.00 are accepted without weakening non-numeric filters
• search evidence follows the mirror's singular/plural tolerance, accepts natural compact-storage spellings such as 128 GB, and recognizes every sort alias implemented by the UI
• product feature predicates check their displayed values rather than treating a present key as truthy; for example, Waterproof: No is rejected
• task 40 no longer requires an invisible women's-only catalog tag; any visibly eligible purple yoga mat is accepted
• adding an already-present target to the cart is graded as an exact quantity increase of one while preserving every other row

Validation

• exercised all 20 accepted task flows in the rendered local UI; Wishlist added only the expected golf polo and Cart added one expected iPhone with a $699.00 subtotal
• 35 environment/verifier tests pass, covering all 20 positive cases plus answer-only shortcuts, wrong-task replay, foreign origins, wrong answers, task-wording plurals, sort aliases, equivalent storage spellings, wrong/missing/non-numeric filters, explicit answer denials, non-eligible and unclicked products, lower-ranked choices, read-only mutations, exact state deltas, other-account state, pre-existing-row preservation/deletion, malformed trajectories/step arrays/invocations, about:blank, and legal alternatives
• all 20 frozen guided UI runs were re-evaluated through agent_demo/eval_judge.py --verifier True against a49afd9 and passed; these are integration replays, not an independent agent success-rate claim
• two isolated code/spec reviews were run on frozen manifests. Both were unpolluted and returned CHANGES_REQUIRED; current head 386ffdbf0869bd5d85adbea690d7a67f89f33287 resolves their complete finding sets with regression coverage
• after task 40's contract changed, it was rerun in the rendered UI using a formerly rejected general-mat alternative; the direct verifier and agent_demo/eval_judge.py --verifier True both passed, and the before/after database snapshots were byte-identical
• Pyright: 0 errors, 0 warnings
• Python byte-compilation and whitespace checks pass
• checked rendered layouts at 320x800, 390x844, 768x900, and 1440x900: no page or top-navigation overflow and no broken images
• exercised homepage to deals, search results, and product details in a real browser
• fetched the asset archive back from pinned HF commit 9c8cb0c112259f07a6d62edfa2f7ee01f94ce3e3 and validated all 1,408 managed members
• candidate archive SHA-256: bcd895293c6837a9f69ccf4e726dee28166005e3089eb939e782eced9768edc1
• source asset SHA-256: cc6e3dafa4d392d35b9113545dd68889ee2bed79e72d5e56e1b1126ffe1f6d31

Integration ordering

Upstream main advanced after this draft was opened and now pins a newer HF asset revision containing the Kaggle and NVIDIA bundles. The only current GitHub merge conflict is .assets-revision. Do not resolve it by replacing the newer global asset line with this older review candidate. Merge HF PR #81 first, then sync the GitHub branch with current main and pin the HF merge commit, which will contain both the newer bundles and the unchanged reviewed Amazon archive.

Not run

A full Docker image rebuild was not run because the host had 38 GiB free, below the 50 GiB safety floor for an unknown-peak full build. Runtime validation reused the existing canonical image with the final site and pinned asset mounted. After the second blind review, only the affected task-40 UI path was rerun; the other corrections are verifier-only and were covered by the 35-test regression matrix.