Add a new domain adapter (CTF, reverse engineering, security audit, scientific literature, etc.)

[jim4226]

The project ships three domain adapters: pr_maintenance (any git repo), self_improve (CSIS itself), and lean_math (Lean formal math, with graceful fallback). A fourth or fifth would meaningfully extend what the system can actually be benchmarked against.

What needs to exist

A new file csis/domains/<name>.py exporting a class that implements three methods:

• graders() -> GraderRegistry — V1 graders specific to the domain
• curiosity() -> Curiosity — frontier-item generator for this domain
• can_run() -> ReadyCheck — graceful fallback (e.g., return ReadyCheck(ready=False, reason='lean CLI not installed') if a dependency is missing)

Domain candidates

Domain	Why interesting
CTF puzzle solving	Concrete benchmark with quantified scoring (each puzzle = 1 point)
Reverse engineering	Builder produces analysis reports; Verifier checks against ground truth
Security audit	Builder produces vulnerability reports; Verifier checks against known CVEs
Scientific literature	Builder summarizes a paper; Verifier checks against the paper's actual abstract
Code review of public PRs	A productized version of pr_maintenance against trending open-source PRs

Acceptance criteria

• New domain module under csis/domains/
• Plumbed into csis/daemon.py:_select_domain() with a CLI flag
• At least one passing regression test
• README + RUN.md updated with the new domain
• Domain-specific graders are pinned (cycle-1 F2 / cycle-2 P3 — pinned source-hash check at cert build)

Reference

• csis/domains/pr_maintenance.py — simplest reference (~100 LOC)
• csis/domains/lean_math.py — example with graceful fallback

Time estimate

~6-12 hours depending on domain complexity.