feat(lwe): add structured recovery challenge

[builtin-pb]

Summary

Adds the Frontier-CS 2.0 task lwe_structured_recovery: recover valid secret vectors for as many as possible of 200 public structured-LWE-style instances. The evaluator is secretless: it reconstructs each public matrix and accepts any submitted secret whose centered residual satisfies the public error and secret predicates.

The PR is intentionally minimal and follows the normal FCS problem layout. It includes only:

• task config/readme/evaluator/evaluate script/reference submission;
• the public catalog and SHA-256 sidecar;
• small public helper modules used by the evaluator and agent-facing solution helper.

Only the files needed for task execution and evaluation are included.

Catalog SHA-256: 318f32b68e8a30c6bcd5867c89eb10d87a5472614401f4e6df0f55356cdc98a6.

Validation

• GitHub Actions: all required PR checks passed after the clean single-commit force-push.
• Catalog loads with 200 instances.
• catalog.sha256 matches catalog.jsonl.
• Public catalog leak scan found 0 non-public metadata fields.
• evaluator.py with the JSON emitted by reference.py returns a clean zero score.
• evaluate.sh with the JSON emitted by reference.py returns a clean zero score.
• Ubuntu 24.04 Docker smoke using CI’s reference.py path returns a clean zero score.
• git diff --check passes.

The branch is one clean commit for review.

[builtin-pb]

LGTM

[joyemang33]

LGTM