[ayu-atp]

This PR contains

Description

This PR adds an Inspect-native sequence-review workflow and fixes cloning
simulator/scorer defects found while reviewing the saved model traces.

Simulator and verifier changes

• Add versioned v2 implementations for PCR, Golden Gate, Gibson, and sticky-end
  restriction assembly while retaining wrappers for the upstream behavior pinned
  at c028ecdcf144b55ffcd92b68be45081df5628c20.
• Recover unique exact PCR amplicons that cross coordinate 0 on circular templates.
• Build multi-fragment Golden Gate products one compatible junction at a time.
  Empty-vector/dropout reassembly remains a plausible candidate; products that
  restore Type IIS sites rank below site-free products.
• Fix Gibson enumeration so a valid circle is not hidden by another possible
  extension, track fragments by input position rather than generated name, and
  deduplicate reverse-complement representations.
• Fix restriction assembly so backbone self-ligation does not suppress a valid
  insert-containing product.
• Score every physically enumerated top-level product rather than result[0].
  Nested multi-product operations fail explicitly as ambiguous because the DSL
  has no product-selection operator. The same candidate must pass both global
  similarity and any digest validator.
• Repair missing reference circularity only when digest metadata proves it during
  normal scoring. The offline rescoring tool also supports the documented
  final-plasmid corpus policy.
• Bump LAB-Bench 2 from 1-A to 2-A because these changes alter cloning scores.

The design rationale, audit findings, and modeling limits are documented in
src/lab_bench_2/CLONING_SIMULATOR_V2.md.

Inspect review tooling

• Render aligned predicted/reference plasmid maps with annotations, readable
  coordinate ticks, topology, digest sites/fragments, sequence differences,
  FASTA views, and recursive sequence provenance.
• Transfer repeated and boundary-modified GenBank features so constructs such as
  the five-copy mCherry reference remain visible.
• Add tools to backfill visualizations and rescore existing traces without
  regenerating model answers. Original scores, explanations, task version, and
  aggregate metrics are preserved in metadata.

Rescored saved submissions

The six *_simulator_v2_reviewed.eval files reuse the recorded answers and include
the rendered comparison event. They have internally updated 2-A aggregate
metrics; the original 1-A summaries remain in metadata.

Run	Original	v2	Changed
cloning_11tasks	3/11	6/11	3
cloning_2tasks_2epochs	0/4	0/4	0
cloning_6tasks_rerun_6M	2/6	3/6	1
cloning_sorcs2_after_session_fix	0/1	0/1	0
gpt56sol_8tasks_method_blind	1/8	6/8	5
gpt56sol_smoke_2tasks_hinted	0/2	2/2	2

Across 32 stored submissions, 11 verdicts change from incorrect to correct. One
method-blind GPT-5.6-sol answer passes using Gibson candidate 3 of 14, directly
showing why the first candidate is not authoritative. The SORCS2 answer now forms
an amplicon but remains incorrect at 0.839761 sequence similarity.

Validation

• make check: all 7 enforced checks pass. Advisory failures are the pre-existing
  private Inspect API import, generated-README dirty-tree heuristic, and a local
  markdownlint Node SSL bootstrap issue.
• LAB-Bench 2 tests: 166 passed and 12 skipped. The two initially sandbox-blocked
  Inspect tests pass when INSPECT_TRACE_FILE is placed under /private/tmp.
• Simulator/scorer/visualization regression tests: 76 passed.

Checklist

• Are you adding a new eval?

  • If yes, please review the Evaluation Checklist.

• Does this change affect existing eval(s)?

  • If yes, have the affected task version(s) been incremented? To check if the version should be incremented, see when to bump the task version.

• Does this change affect how future contributors write or submit evaluations (e.g. new required fields, changed tooling, updated conventions)?

  • If yes, has the relevant documentation been updated (e.g. CONTRIBUTING.md, EVALUATION_CHECKLIST.md, AGENTS.md)?