`/hunt` jurisdiction system: configurable LLM-as-judge with strict/evidentiary/permissive policies

[Hmbown]

Full jurisdiction system: configurable judge LLM, three policies, trajectory-aware verdicts.

After #2092 (vocabulary, v0.8.45) and #2093 (verifier-preview wiring, v0.8.46), this issue lands the full Codex-style LLM-as-judge with codewhale-native jurisdictions. Absorbs and supersedes the original "port Codex goal system" framing.

Concept

A jurisdiction is the policy a judge applies to decide whether a quarry is statutorily hunted. Three built-in jurisdictions:

Jurisdiction	What counts as hunted
strict	Diff exists, tests added or updated, CI green (or local verifier equivalent).
evidentiary	Diff cites files; agent shows changes; no contradiction with quarry.
permissive	Agent declares done; judge sanity-checks.

Each turn ends with a judge call. Judge sees: quarry, current trajectory summary, this turn's evidence, full diff against origin/main (or session start). Judge returns:

struct JudgeRuling {
    verdict: HuntVerdict,
    reasoning: String,
    next_step: NextStep,  // continue | handoff | abandon | declare_hunted
}

In scope (v0.9.0)

• config.toml [hunt] section: jurisdiction = "evidentiary" (default), judge_model = "auto" (defaults to the session model with a judge-only system prompt), judge_max_tokens = 4096, judge_temperature = 0.0.
• Judge runs at each turn boundary in hunt mode. Verdict logged inline in transcript between turns.
• next_step actions:
  • continue — model gets a system-message hint of judge reasoning, continues.
  • handoff — runtime suggests a sub-agent and lists the species + brief.
  • abandon — session marked escaped, no trophy.
  • declare_hunted — trophy written, session may auto-close per config.
• /hunt jurisdiction <strict|evidentiary|permissive> switches mid-hunt.
• /why hunted / /why wounded / /why escaped shows the judge's last reasoning.
• Judge prompt template under crates/tui/src/prompts/judge.txt — small, focused, source-controlled.

Out of scope

• Trained verifier models (out of scope forever for v0.9.0; remains LLM-as-judge).
• Per-statute custom jurisdictions in config (RFC; lands later if asked for).
• Judge-as-sub-agent species (the judge isn't a whale — it's the court).

Acceptance

• Three jurisdictions selectable and observably different in behavior on the same test quarry.
• Judge prompt is auditable and source-controlled.
• Verdict + reasoning render inline in transcript.
• /why <verdict> returns the most recent judge reasoning.
• Trophy card includes the jurisdiction the hunt was decided under.
• Eval harness has at least one regression test per jurisdiction asserting the verdict shape.

Closes / partially closes

• Closes v0.9.0: Port Codex goal system — LLM-as-judge with persistent continuation loop #2058 (Port Codex goal system) — this is the codewhale-native realization, not a port.
• Builds on Rename /goal → /hunt: quarry, verdict vocabulary, trophy card writer #2092 (vocabulary) and Wire verifier preview to emit hunt verdicts (pass/partial/fail → hunted/wounded/escaped) #2093 (verifier wiring).

Notes

• The judge isn't a sub-agent species. Whales hunt; the judge is the court. Treat as a distinct primitive.
• LLM-as-judge has known failure modes (sycophancy toward the agent, brittleness on edge cases). The three-jurisdiction split is partial mitigation: strict is mechanical (CI gates aren't LLM-decided), permissive is honest about its laxness, evidentiary is where the judging actually happens.

Replaces #2058. Final piece of the hunt trilogy: #2092 (vocabulary, v0.8.45) → #2093 (verifier wiring, v0.8.46) → this (v0.9.0).