Review: Amtrak mirror + task verifiers (site by @Lxr-max, verifiers by reviewer)

[JeremyJC67]

This PR is the review vehicle for #39. It keeps @Lxr-max's two site commits intact
(authorship preserved, rebased onto current main so the PR is mergeable) and adds
the reviewer deliverable: a deterministic grading contract for every task, validated
against the running mirror. Full review with evidence is posted on #39.

What is in here

Commit	Author	Content
2af6a47 Add Amtrak WebHarbor mirror	@Lxr-max	the site (sites/amtrak/), registration as the 25th site on port 40024 (rebased; the PR's stale 15/16-site doc, skill and .gitignore edits were dropped during the rebase because main had moved on)
d8d964a Finalize Amtrak asset workflow and checks	@Lxr-max	contributor follow-up (seed / template polish)
4f18dce chore(amtrak): sweep docs and tests to 25 sites	reviewer	README / AGENTS / CONTRIBUTING / CLAUDE / agent_demo README to 40000-40024; walmart_careers + rotten_tomatoes registry tests updated; tasks.jsonl web -> 40024 (port provisional until merge)
fcd4697 test(amtrak): verifiers, verify_lib, validation harness	reviewer	sites/amtrak/verify/verify_lib.py, verify_0.py ... verify_17.py, verify/README.md, verify/tests/ (offline unittest matrix + Playwright live matrix, excluded from the image by .dockerignore)
57c92ae chore(amtrak): grading contract in tasks.jsonl	reviewer	verifier_path + judge_rubric on all 18 rows; tasks 2, 3, 5 re-anchored, 16 replaced, 8 others clarified (no answer key anywhere)

Grading contract

Deterministic-first, no LLM dependency (--no_llm True changes nothing):

1. package validation: task id, non-empty answer, completed run, loopback single-origin URLs, decodable PNGs;
2. navigation gates: /booking/results query-parameter gates (station aliases, date, sort, fare class, passengers), ordered funnels (round trip, multi-city, sleeper rooms, checkout), login as the named account;
3. answer matchers: affirmative money / duration / date / booking-code / track / stop-order matchers against hardcoded ground truth;
4. SQLite snapshot contract: 22-table schema, seeded counts, immutable catalog; read-only tasks require every mutable table unchanged; tasks 8 and 17 require exact row deltas (preferred-station change; one new Business booking with its segment, ticket, passenger, mock payment, reward activity and points credit). search_logs is excluded because the site writes it on every search / results request.

Validation

• Offline matrix (python -m unittest discover -s sites/amtrak/verify/tests): 13 tests, all pass (genuine, no-op, shortcut, wrong answer, other task id, unterminated, foreign origin, corrupt PNG, schema tamper, collateral write, state mismatch, over-reach, phrasing variants).
• Live matrix (verify/tests/live_matrix.py, real Chromium against the standalone mirror): 18 tasks x {no-op, pass, shortcut, wrong, state-mismatch} = 90 cells, all as expected (every no-op / shortcut / wrong / state-mismatch run FAILs on the intended check; every genuine run PASSes).
• agent_demo/eval_judge.py --run_dir <run> --verifier True returns exit 0 on the genuine runs and exit 1 on the no-op.
• Standalone mechanical checks: py_compile ok, asset archive validated (308 members, no junk), seed md5 cb4b6a5939b663017d9774f2d3a86884 byte-identical after first and second boot, no cross-site imports, all 215 internal links 200, no broken images or page errors.

Still required from the contributor (details on #39)

1. Assets: amtrak.tar.gz is not in HF discussion Add NBA mirror #30 (refs/pr/30 holds only the 15 base tarballs); it exists only on the fork commit lxr-max/WebHarbor@eac108f. Open an HF PR that actually contains the tarball, merge it, then repin .assets-revision (currently revision: main; main is pinned to 65c479f8…).
2. account() "Next trips" lists Completed trips first — filter by is_upcoming.
3. /search, /help?q= and /booking/results commit search_logs rows, so read-only browsing is not byte-identical (the reset invariant still holds; the verifiers tolerate it).

Full 25-site Docker environment built from this branch with amtrak.tar.gz extracted by hand from lxr-max/WebHarbor@eac108f (no ChilleD/WebHarbor revision carries it yet): build ok (4.70 GB), 25/25 sites 200, /health 25/25, POST /reset/amtrak byte-identical (cb4b6a59…) before/after a dirty request and after docker restart, /reset-all 25/25 in 1.64 s, verifier no-op matrix in the container 18/18 FAIL, registry tests green, offline verifier suite 13 passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AkNTAZMziDodR6uqa6QBtP

[JeremyJC67]

Rebased onto the 26-site main — this branch is now site 27 / port 40026

main merged FedEx (40024) and WebMD Doctor (40025) overnight, which put every open site
branch into conflict. This branch has been brought up to date by merging main rather than
rebasing, so every contributor and reviewer commit is preserved verbatim, and re-slotted to
index 26 / port 40026: registry, Dockerfile EXPOSE, the shared docs and every
site-local port reference follow. The slot is provisional as always — say the word at merge
time and I will re-slot in whatever order you land things.

The whole environment was then rebuilt and re-verified at the new slot:

check	result
HTTP sweep	27/27 return 200
GET /health	ok, 27/27 alive and ready
POST /reset/amtrak → md5	cb4b6a5939b663017d9774f2d3a86884, identical to the seed
POST /reset-all	ok, 27/27 ready, 1.50 s
verifier no-op matrix	18/18 FAIL
offline verifier suite in the image	15 passed, 119 subtests

Audit assertions hold: GET /logout 405, branded 404 (10,664 bytes), and the input validation
that replaced the silent fallbacks — an empty booking search, origin=a, and
passengers=100000 all return 400. Three read-only page loads leave the DB byte-identical,
which is the log_search() fix confirmed in the image.

The asset blocker is unchanged and is the only thing left: no ChilleD/WebHarbor revision
contains amtrak.tar.gz. It needs an HF PR against the central dataset before this can merge.

[JeremyJC67]

The asset blocker on this PR is now open rather than absent — I opened it on @Lxr-max's behalf, since it was the one remaining item that no maintainer could act on and the site author has not been able to.

HF PR #77 adds amtrak.tar.gz to the central dataset. The bytes are taken verbatim from @Lxr-max's own fork at lxr-max/WebHarbor@eac108f14be667b7e78972b0805d116ddea931d2 — nothing was repacked, re-rendered or modified, and the description credits them as the site author:

116,525 bytes   md5 1224b94239a4e23ace5691f632c896ee
308 managed members, rooted at amtrak/, 0 AppleDouble or .DS_Store entries
scripts/validate_asset_archive.py -> validated 308 managed members for amtrak
instance_seed/amtrak.db md5 cb4b6a5939b663017d9774f2d3a86884
  (the same hash this review verified against, and the same one the container run reproduced)

@Lxr-max — if you would rather open this yourself, say so and I will close #77; the point was only to unblock the queue, not to take the step from you.

With that, everything on this PR that was not a maintainer action is done: the branch has merged the 26-site main, re-slotted to port 40026, and passed a full 27-site container run (27/27 sites 200, reset byte-identical cb4b6a59…, reset-all 27/27 in 1.50 s, verifier no-op matrix 18/18 FAIL). The remaining sequence is: merge HF #77, then I pin .assets-revision to the merge sha the same day, exactly as I have just done for Healthline now that HF #70 landed.