Review: Add Healthline mirror + task verifiers (site by @JeremyJC67, verifiers by reviewer) (#59)

[TabsPhasers]

Summary

基于当前 main 接管 @DEM1TASSE 的 #59（Healthline 镜像 + task verifiers）。本 PR 为 Draft 占位版本，用于承载 reviewer 的修复与验证结果。

⚠️ Draft / not mergeable yet：本分支尚未重铺到当前 main，GitHub 显示 CONFLICTING（mergeable=false / dirty）。当前 diff = 原贡献 + 本次 reviewer 修复，共 53 个文件。下一次 push 会重铺到当前 main 并解决冲突，diff 收敛为 sites/healthline/** + Dockerfile / control_server.py / websyn_start.sh + tasks.jsonl / verify/*。在重铺与素材合并完成前，请勿将本 PR 作为合并候选。

Original review resolution

1. 端口修正 40015→40016 仍与 main 冲突：当前 main 的 40016 已归 ikea；改为追加式分配 40024，并保留 main 全部站点（25 站）。
2. Healthline--3/16 rubric 已补全：对应 verifier 精度不足——离线探针发现 11 个确定性假阳性（数值子串、复合症状重复计数、否定句通过、药物类别颠倒、缺 DB 当 0 等），已全部修复。
3. .assets-revision：素材 HF PR Add AKC mirror site #40 未合并、原 pin 为 refs/pr/40；本轮不改 pin（保留 main 的稳定 revision），待 HF Add AKC mirror site #40 合并后 pin 到 merge SHA。

Review changes

• sites/healthline/verify/* + verify_lib.py：数值边界匹配、否定/肯定区分、药物-类别关联、DB 缺失 fail-close、导航门收紧、分组计数。
• sites/healthline/tasks.jsonl：web 端口 40016 → 40024。
• Dockerfile / control_server.py / websyn_start.sh：保留 main 站点列表并追加 healthline（端口 40024）。

Validation

• Baseline no-op：20/20 正常 FAIL（无假阳性放行）。
• 纯视觉原生 actor：20/20 真实轨迹（含 5 次宿主 watchdog infra 重试闭合）。
• 修复版确定性 verifier（--no_llm）：20/20 exit 0。
• 修复版 verifier（LLM 检查开启）：20/20 exit 0。
• 独立 LLM judge（官方 eval_judge.py 语义，轨迹 + 末 4 帧）：20/20 success，无 false checkpoint。
• 主审逐题读图 + before/after DB + trajectory：20/20 PASS。
• 完整证据包（per-task 表、原始输出、哈希）保存在 reviewer 本地目录，不在本 PR 提交中。

Asset delivery status

• HF PR：https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/40（Open）
• .assets-revision：保持 main 的 revision（本 PR 未修改）。在该 PR 合并并 pin merge SHA、clean-fetch 通过前，本 GitHub PR 不得合并。

Scope

• 保留原作者 @DEM1TASSE 及原站点贡献归属。
• 排除 reviewer 运行产物、截图、缓存与本地证据。
• 共享框架改动：无。

Blocking checklist

• 重铺到当前 main 并解决冲突，把 diff 收敛为站点 + 必要集成改动
• HF PR Add AKC mirror site #40 合并 → .assets-revision pin merge SHA → clean checkout 执行 fetch_assets.sh healthline + build/reset 验证
• 按 reviewer 口径补登记接管说明（本项目此前已声明跳过 lock comment）

Reviews #59.

[JeremyJC67]

@TabsPhasers thanks for picking this up. A heads-up so you don't redo work that is already
done, and an offer at the end.

This branch is based on #59's July head (ports 40015/40016). The site PR #53 was rebased onto
current main on 2026-09-11: healthline is site 25 / index 24 / port 40024, @DEM1TASSE's two
verifier commits are cherry-picked in with their authorship preserved, and it is MERGEABLE/CLEAN.
So the first item on your blocking checklist is already closed there.

It has since had a second pass against the maintainer merge-blocker checklist, verified on a full
25-site container build: 25/25 sites 200, reset and restart byte-identical (a8a9b54…), reset-all
1.50 s, no-op verifier matrix 20/20 FAIL. Two of those findings were blockers rather than polish:

• a hardcoded SECRET_KEY — signing {"_user_id":"1"} with the in-repo key and requesting
  /account returned 200 rendering alice.j@test.com;
• GET /article/<slug> incremented view_count and committed, so read-only browsing changed the
  DB md5 and the 15 read-only tasks could not assert tables_unchanged.

On assets: HF #40 is closed. It was superseded by
#70, same bytes rebased on
current asset main. Keeping .assets-revision at main's revision will not work either — no
merged revision contains healthline.tar.gz yet, so fetch_assets.sh healthline fails until #70
merges. #53 pins refs/pr/70 temporarily and documents the repin, the same way #86 and #95 do.

Three of your verifier predicates are things I do not have, and they are good: count_groups
in verify_3 (so overlapping symptom tokens count once), classes_not_reversed in verify_16 (which
rejects swapping the two drug classes), and amount_with_unit. Your boundary-aware
number_mentioned matches a fix I landed independently — useful corroboration that it was a real
hole.

So rather than two vehicles for one site, either of these works and both credit your review:

1. you rebase this PR onto feat(healthline): add Healthline mirror + task verifiers (site 25, port 40024) #53's head and it stays the vehicle; or
2. you review feat(healthline): add Healthline mirror + task verifiers (site 25, port 40024) #53 directly, and I cherry-pick those three predicates from cdf5865 with your
   authorship preserved.

Tell me which you prefer and I will do my side of it today.

One small correction for the record: the Scope section credits @DEM1TASSE as the original author.
They authored the verifiers; the site itself is mine — which your title already states correctly.

[TabsPhasers]

Thanks — we'll take option 2: #53 stays the merge vehicle. We're closing #105 as superseded; the branch remains at TabsPhasers/WebHarbor:review/pr-59-healthline @ cdf5865 if you want to diff anything.

Requested cherry-pick with authorship preserved — 8 predicate deltas, not just the three helpers you named (we diffed our tree against #53's head 79e03ce):

1. verify_lib.count_groups + verify_3 — overlapping symptom tokens ("slow-healing" + "sores", "thirst" + "urination") must count once; feat(healthline): add Healthline mirror + task verifiers (site 25, port 40024) #53 still uses count_matches, so a one-symptom answer can pass as three.
2. verify_lib.amount_with_unit + verify_11 — require a standalone 50 mg with a unit; substring matching still accepts "150 mg".
3. verify_16.answer_classes_correct (classes_not_reversed) — reject swapping the two drug classes; contains_all alone passes "lisinopril is a statin; atorvastatin is an ACE inhibitor".
4. verify_6 — fail-close when the initial DB is unavailable (saved_articles_for(...) or [] currently coerces "no DB" into "0 saved", which a wrong answer can satisfy).
5. verify_8.answer_reviewer — require the reviewer's name (Kim Chin), not just the credential string.
6. verify_15.answer_affirms — broadened negation/affirmation set; the current four-phrase list passes "No, grapefruit is absent."
7. verify_17.answer_condition — require the condition identity (hypertension / silent killer), not only the detection method.
8. verify_19 — same grouping fix as (1) for migraine triggers ("bright light" + "light" double-count, etc.).

number_mentioned no longer needs taking from us — your independent anchoring is equivalent to ours; treat it as corroboration only.

Two things for the record:

• Your two blockers (hardcoded SECRET_KEY; GET /article/<slug> committing view_count) are real gaps our pass missed. We're recording them as findings on our side; thanks for catching them.
• Scope credit is as you corrected: site by @JeremyJC67, verifiers by @DEM1TASSE; our reviewer deltas are the 8 above. We've fixed our PR text accordingly.

On assets we're aligned: HF #40 closed → #70; .assets-revision temporary refs/pr/70 is the same pattern as #86/#95, and the one-line repin to the merge SHA follows once #70 lands. We'll re-verify the predicates after your cherry-pick lands and can post a short reviewer confirmation on #53 if useful.

If a direct cherry-pick conflicts, we're happy to push a small patch branch on top of #53's head instead — just say which you prefer.

(This reply was posted automatically via OpenCode.)

[Raibows]

Thanks for your contribution! @TabsPhasers @JeremyJC67 @DEM1TASSE

[JeremyJC67]

@Raibows @TabsPhasers — congratulations on landing this, and one follow-up finding on the merged
tree, since it is the single item from the #53 audit that did not come across.

Two unvalidated redirects are live in sites/healthline/app.py on main:

app.py:700   return redirect(request.args.get("next") or url_for("account"))
app.py:846   return redirect(request.referrer or url_for("article_detail", slug=art.slug))

Both follow an attacker-supplied absolute URL. \/login?next=https://evil.example/pwn lands the
signed-in user off-site, and the second does the same through a forged Referer. Everything else
from that audit did make it — HEALTHLINE_SECRET_KEY with a random fallback, the read-only
GET /article/<slug> (the docstring on article_detail states the invariant), MAX_CONTENT_LENGTH,
the branded error handlers, count_groups / amount_with_unit / tables_unchanged — so this looks
like an oversight in the integration rather than a decision.

The fix that was on #53 is small and self-contained:

def safe_redirect_target(target, fallback):
    """Only follow same-origin relative paths; anything else falls back."""
    if not target:
        return fallback
    parsed = urlparse(target)
    if parsed.scheme or parsed.netloc or not target.startswith("/") or target.startswith("//"):
        return fallback
    return target

with the two call sites becoming redirect(safe_redirect_target(request.args.get("next"), url_for("account")))
and the equivalent for the Referer case.

Say the word and I will open a one-file PR against main with exactly that, plus a container run to
confirm both redirects fall back. It is worth noting that the same pattern is likely present on other
merged sites — the audit found it on several — so if you would rather have it as one sweep across the
registry than a per-site fix, that works too.