feat(eve): accept per-selection reasoning in dynamic model selection (#629)

[ZhYGuoL]

A defineDynamic selection accepted { model, modelContextWindowTokens, modelOptions } but not reasoning, so a downshifted session kept the agent-level reasoning (e.g. xhigh), defeating the point of dropping to a cheap classifier.

What changed

• Add reasoning?: AgentReasoningDefinition to PublicAgentModelSelectionDefinition and to the runtime model reference InternalAgentModelDefinition (packages/eve/src/shared/agent-definition.ts).
• Add reasoning to DYNAMIC_MODEL_SELECTION_KEYS and validate it in validateDynamicModelSelection against the same enum the agent-level field accepts, reusing the now-exported normalizeAgentReasoningDefinition (packages/eve/src/runtime/agent/resolve-model.ts, packages/eve/src/internal/authored-definition/core.ts). Both selection-shape error strings (resolve-model.ts:192 and :236) now list reasoning?.
• Carry the value onto the reference in normalizeDynamicRuntimeModelResult, and apply it at the model-call precedence point: session.agent.modelReference.reasoning ?? session.agent.reasoning (packages/eve/src/harness/tool-loop.ts:852). Per-selection overrides the agent default; omitted falls back to it. Same precedence as modelContextWindowTokens/modelOptions.
• Document the field on the selection object (docs/agent-config.md).

How I verified

• Before: a resolver returning { model, reasoning } throws unknown key(s): reasoning at resolve-model.ts:178. Reverting only the tool-loop.ts precedence line, the override test sees the model call get reasoning: "high" (agent default) instead of "low" (per-selection).
• After: per-selection reasoning lands on the reference and reaches the model call; omitting it keeps the agent default. Invalid values are rejected with a reasoning-specific error.
• Tests (packages/eve/src/runtime/agent/resolve-model.test.ts, packages/eve/src/harness/tool-loop.test.ts): "carries a per-selection reasoning onto the reference", "leaves reasoning unset when the selection omits it", "rejects a selection with an invalid reasoning value", "lets a per-selection reasoning override the agent reasoning effort". Each fails before, passes after.
• Full local green: pnpm lint, pnpm typecheck, pnpm test:unit (4687 passed, 1 skipped), pnpm docs:check.

Problem 2 from the issue (small-model outputSchema reliability) is out of scope for this PR.

Closes #629.

[vercel]

@ZhYGuoL is attempting to deploy a commit to the Vercel Team on Vercel.

A member of the Team first needs to authorize it.