Register `--prompts soul` variant reading from `SOUL_MD_PATH`

[EYH0602]

Context

SOULFuzz is a benchmark-agnostic
prompt-optimization framework. It iteratively mutates a SOUL.md artifact (a
markdown file injected as system context) and evaluates each candidate against
a pluggable benchmark. KoraBench is the first adapter, used for the
child-safety case study.

To close the optimization loop, the SOULFuzz Python adapter needs KoraBench to
read the system-prompt body from a file path the adapter controls per
invocation. This is touch point #2 of the four-edit adapter contract
described in the SOULFuzz engineering plan §6 — minimal, additive, and
intended to be upstreamable.

Ask

Register a new --prompts soul variant in the run command, alongside the
existing default / child variants. Behavior:

1. When invoked with --prompts soul, KoraBench reads the system-prompt body
   from process.env.SOUL_MD_PATH and uses that as the system prompt for the
   target model.
2. If SOUL_MD_PATH is unset or the file is missing, fall back to
   data/souls/seed.md resolved relative to the KoraBench checkout root. If
   that file is also missing, exit with a clear error (not a silent fallback
   to default).
3. Read-once-per-run: load the file when the prompt variant is resolved, not
   per-scenario. The file does not change during a single yarn kora run
   invocation.
4. No changes to the JSON output schema, judge plumbing, or scenarios are
   required. This is strictly a new entry in the prompt-variant registry.

Acceptance criteria

• yarn kora run <target> --prompts soul --limit 1 succeeds when
  SOUL_MD_PATH points at a readable markdown file, and the file content
  appears as the system prompt in the resulting transcript.
• Same invocation with SOUL_MD_PATH unset and data/souls/seed.md
  present uses the seed file.
• Same invocation with both unset and missing exits non-zero with a
  message naming SOUL_MD_PATH.
• A unit test covers the three branches above (env set / env unset with
  seed / both missing).
• No regression to --prompts default or --prompts child.

Out of scope (separate issues / PRs)

• HCB judge registration (touch point Register --prompts soul variant reading from SOUL_MD_PATH #4 from the SOULFuzz engineering plan).
• custom-cfa-* model slugs and the CFA workspace (touch points docs: align README node requirement with .nvmrc (24 LTS) #1 + the
  packages/cfa/ workspace).
• Optional intimacyCalibration mechanism (touch point [feat] register openai as built-in openai-compatible provider #3).

References

• SOULFuzz repo: https://github.com/SecurityLab-UCD/SOULFuzz
• Tracking PR-03 on SOULFuzz side: lands the Python adapter
  (Evaluator + Splits + Injector + seed) and bumps this fork's
  submodule pin once this issue is resolved.
• The seed for data/souls/seed.md is a verbatim copy of the existing
  --prompts child body — the SOULFuzz adapter will produce it at bootstrap
  time, so no new content needs to be committed in this fork.