Integrate Amtrak PRs #39 and #99 with reviewed UI, task, and grading fixes

[QianhuiWu]

Amtrak: PR #39 → PR #99 → reviewed fixes

Integration date: 2026-09-17. Original site contribution: @Lxr-max.
Review continuation: @JeremyJC67. Follow-up fixes are based on PR #99.

History and scope

Regular merge commits preserve original PR #39 head
ab903781556b72f6900c7be2d89c89cc46a6c0d9, review PR #99 head
677b87c007bcf7390a6dea51171033e75681bdf6, and fix commit a57ad6a.
No squash, rebase, or force-push is used for this integration. The original and
review branches both remain ancestors even though the reviewer had previously
rebased/copied the contributor's commits.

Amtrak is appended at index 39, port 40039. All existing 39 site ports and
site implementations are preserved, along with main's seed migrations and
asset pins. The integration is isolated from the user's running previews.

Reviewed fixes

• Restore the homepage multi-city link and usable mobile booking/search layout.
• Update fare totals immediately and include checkout fees and room prices;
  preserve selected fares and reject invalid fare submissions.
• Strengthen ten short tasks (5, 6, 7, 9, 10, 11, 13, 14, 15, 16) with substantive
  comparisons, constraints and multi-page evidence, not click padding.
• Use natural instructions and responses for all 18 tasks; remove mandatory
  JSON schemas, internal field names and exact-policy-quotation requirements.
• Align rubrics, bounded natural-answer grading, immutable fixture checks and
  precise saved-state verification. Keep structured-answer backward compatibility.
• Record explicit before/after page text and post-action URLs while preserving
  main's existing pre-action observed_text and final-observation fields.

Merged HF assets

HF PR #96
publishes the original contributor's unchanged amtrak.tar.gz into
ChilleD/WebHarbor, at merged commit
6f432484307f76ca3c58b1d5c8f7b614dfb34f54.
Incomplete draft #30 contained no Amtrak archive and was closed with a link
to the replacement. All 44 previously existing dataset files are unchanged.

• Archive: 308 validated managed members.
• Archive SHA-256: 22f0132a5f9bfee45a641e8172050a94bb4fc0de4ad45f1d9e52deb50eb37f34.
• Seed SHA-256: b245928c8b03741d17aa7c27f863a499e44034af607902cfa9c8d099c93591b4.
• The scoped central pin and digest are tracked in .assets-revision.
• All 40 sites were freshly fetched and extracted in the isolated integration
  checkout; assets were not copied manually from the preview.

Verification

• 11 asset-fetch tests pass, including preservation of existing scoped pins,
  explicit overrides, missing-archive behavior and digest rejection.
• 16 Amtrak unittest methods pass, including natural paraphrases, adversarial
  claims, package/fixture/state checks and explicit before/after URL pairing.
• The official evaluator regraded 182 preserved evidence packages: all matched
  expectations (18 prior browser recordings and 164 synthetic controls).
  These regrades are not new browser runs.
• Python compilation, shell syntax, whitespace and the 40-site registry pass.
• Combined Docker build succeeds: webharbor:pr99-integrated, image
  sha256:978abcae6f8464b5f8ff31c5ee4138090b8f4cb7053a30c36ac60ef982843758.
  All 40 sites are healthy and all 40 homepages return HTTP 200.
• Four fresh scripted Chromium regressions pass the official evaluator:
  task 3 (multi-city), 8 (saved profile), 16 (refund/fare comparison), and
  17 (completed Business booking). These are scripted checks, not LLM-agent runs.
• Reactive fare totals change from $69.11 to $88.53 for the tested selection.
  Mobile homepage has no horizontal overflow; booking panel width is 318 px
  at a 390 px viewport. Integrated homepage matches the reviewed preview.
• Actual profile mutation changes the runtime DB; reset and service restart
  both restore byte identity with the seed:
  cb4b6a5939b663017d9774f2d3a86884 (MD5).
• Shipped Amtrak source hashes match the tested worktree; verifier code compiles
  under container Python 3.12. Runtime source candidate: b6e9ff1; this report
  is the only subsequent tracked addition.

Local integration evidence: .assets/integration-pr99/ in the integration
worktree. Earlier task GIFs remain in the fix worktree at
.assets/reviews/pr99-fix-20260917/natural-answers/.
The existing dashboard is http://localhost:43815/natural-answers/ and the
reviewed standalone preview is http://localhost:44815/ (forward both ports
when accessing the remote workspace). They are not deployed release URLs.

Limits

The mirror uses synthetic schedules, fares and generated artwork, not live
Amtrak data or a faithful copy of every upstream page. The natural-answer parser
has bounded coverage, not unrestricted semantic understanding. No independent
LLM-agent run or secondary LLM-judge run is claimed. Docker Hub publication and
deployment are separate stages, not included in this source/assets merge.
The combined image is 5,342,471,387 bytes (5.34 GB), above the repository's 4 GB
target. Broad inherited image-size cleanup was not part of this integration.

[QianhuiWu]

Integration verified complete: main is now 454e7a4, with a tree identical to the validated PR head. GitHub reports original #39, review #99, and this PR all MERGED; their actual ancestry is preserved in that order. Required HF PR #96 is merged and pinned; incomplete HF draft #30 is closed with a replacement link. Fresh 40-site fetch/build, all 40 health/homepage checks, 182 expected grading results, four new scripted browser regressions, and byte-identical Amtrak reset/restart passed. Local main is fast-forwarded and clean. The temporary test container is stopped; existing previews remain running. Full report: review-reports/amtrak-pr99-integration.md. Docker Hub publication/deployment was not performed; the local image is 5.34 GB.