Review: Add FedEx WebHarbor mirror (original by @Lxr-max)

[jackjin1997]

Summary

Reviewer-owned completion PR for #50, preserving and attributing the original FedEx contribution by @Lxr-max while integrating it with current main.

This follow-up:

• fixes the FedEx environment blockers found during review, including invalid SVG assets, malformed numeric input handling, and the missing create/remove workflow;
• updates the 18-task contract and adds one deterministic verifier plus one non-answer-leaking judge rubric per accepted task;
• pins the exact reviewed HF asset candidate from https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/50;
• retains the original contributor commits in the branch history.

Verification so far

• Official repository image build: passed — https://github.com/jackjin1997/WebHarbor/actions/runs/32224183751
• Full environment smoke: 17/17 HTTP and health checks passed before and after reset-all; FedEx reset restored the seed DB byte-identically.
• Browser QA: 19/19 environment scenarios and 18/18 task-contract scenarios passed in real Chromium.
• FedEx tests: 15/15 passed.
• Recorded task executions: 18/18 canonical runs passed deterministic verification.
• Adversarial verification: 65/65 no-op, shortcut, wrong-answer, contradictory, extra-entity, and state-mismatch cases were rejected.
• Reviewed asset: archive SHA-256 c34590c011bf1d8cda175ee2232934d5a2228c271199b404a9219c3056ba4f05; 25 members; no AppleDouble metadata; seed DB MD5 ab9cb498afbd61bb1955c6d63aa24642.

Status

This PR intentionally remains Draft while the required independent blind task review and final reconciliation run. The paired HF PR may remain open during review, but it must be merged first and .assets-revision updated to the resulting dataset main SHA before this code PR is merged.

Original PR: #50

[jackjin1997]

Independent blind task review

A fresh read-only Claude Code 2.1.235 / Opus 5 pass reviewed the frozen 18-task handoff from raw tasks, rubrics, trajectories, screenshots, and before/after artifact hashes. It had no verifier verdicts, prior review conclusions, repository edit access, or web access.

Result: 18 PASS / 0 FAIL.

Four non-blocking observations are being reconciled before the Draft is promoted: clarify one Dallas task phrase, directly confirm the two mutation rows in SQLite, confirm the default Seattle pickup location was submitted, and explain expected query-log DB deltas on two search tasks. I will post the reconciled result and final verification after those checks.

[jackjin1997]

Reconciled final verification

All reviewer gates pass at ad0ec98, and this Review PR is Ready for final maintainer review.

• Latest complete CI: https://github.com/jackjin1997/WebHarbor/actions/runs/32229641337 — exact reviewed HF assets, 15/15 FedEx tests, official scripts/build.sh, 17/17 HTTP and health before/after reset-all, byte-identical FedEx reset, reset-all in 2s.
• Task contract: 18 accepted tasks, 18 deterministic verifiers, 18 non-answer-leaking rubrics.
• Real task runs: 18/18 canonical PASS; adversarial suite rejected 65/65 invalid cases.
• Independent blind review: Claude Code / Opus 5 returned 18 PASS / 0 FAIL.
• Reconciliation: the Dallas invoice task was clarified test-first to make destination direction explicit, then rerun in real Chromium, re-verified, and independently re-reviewed PASS.
• Final review state: clean worktree, git diff --check PASS, current upstream main included, GitHub reports the head mergeable and clean.

The paired HF PR remains open at the exact reviewed commit c147b46db6f182b9bd29d2189f93d3808a409028: https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/50. Its archive SHA-256 is c34590c011bf1d8cda175ee2232934d5a2228c271199b404a9219c3056ba4f05.

This completes reviewer preparation. Final approval, merge order, merging, and any post-merge pin decision belong to the maintainers; I will respond only to requested review changes.