feat(one-reason): add explicit Jev routing with LLM compatibility

[yaonyan]

Summary

Adds optional TypeSafe Jev decision routing to @one-agent/reason while preserving the historical LLM default.

Final behavior

• Omitted mode or mode: "llm": always uses the existing LLM path.
• mode: "jev": decision-shaped examples use Jev; free-text examples or unavailable Jev fall back to LLM.
• auto is not a public mode.
• Jev credentials configure availability only; they do not change the default mode.
• ONE_REASON_JEV_ENABLED=0 disables Jev availability.

Example schema mapping

• boolean -> Jev noul / Gateway boolean
• number -> Jev score
• plain strings remain free text and use LLM fallback in Jev mode
• arrays remain arrays
• choices must be explicit:

{
  route: {
    $jev: "choice",
    options: ["approve", "manual", "deny"],
  },
}

Score ranges can be declared explicitly:

{
  risk: {
    $jev: "score",
    range: [0, 100],
  },
}

The Jev request uses at most 10 levels and maps the returned score back into the declared range. Invalid ranges and more than 10 custom criteria fail locally before a network request.

Providers

• TypeSafe: POST {baseURL}/systemone
• AI Gateway: POST {baseURL}/evaluation-model
• Provider credentials are independently configurable from the primary LLM.

Verification

• TypeScript typecheck
• 40 unit tests
• package build
• real TypeSafe Jev smoke test
• real AI Gateway Jev smoke test
• real LLM routing smoke test (request reached the provider; OpenRouter returned HTTP 402 due to insufficient credits)
• local score range/criteria validation tests

[yaonyan]

Hybrid Jev / LLM fallback

Addressed the feedback that summary-style reason() is useless with Jev (Jev only returns noul/choice/score).

Product story: Jev = control-node judgments; summary/synthesis stays on LLM (auto-fallback under typesafe/gateway).

Changes (915f940):

• reason(prompt, example, options?) with mode: "auto" | "jev" | "llm", optional questions / state / booleanThreshold
• auto (default): decision-shaped → Jev; free-text/summary → LLM fallback
• LLM fallback via ONE_REASON_FALLBACK_* or one auth model
• README documents decision vs summary examples
• Tests: decision→Jev, summary→fallback, all three modes (35 passing) + typecheck OK