[BUG]: agent-level `request.body` and `request.settings` are populated from config and never reach the provider

An agent carries a request field (packages/schema/src/agent.ts:27, request: Provider.Request),
where Provider.Request is { settings, headers, body } (packages/schema/src/provider.ts:55-59).
Config populates it, and that population is covered by a test. Nothing on the request path ever
reads it, so per-agent generation settings are silently ignored.

Writer, packages/core/src/config/plugin/agent.ts:112-113:

Object.assign(agent.request.headers, item.request.headers ?? {})
Object.assign(agent.request.body, item.request.body ?? {})

Covered by test, packages/core/test/config/agent.test.ts:129:

expect(agent.request).toEqual({ settings: {}, headers: {}, body: { temperature: 0.5 } })

Never read. packages/core/src/session/model-request.ts receives the agent as Agent.Info
(:74) and reads exactly one field off it — input.agent.system (:84-85) — plus input.agent
as scope identity (:206). Generation options have a single source: the draft is seeded empty at
:217 (options: {}) and filled only by whatever a session context/generate hook writes,
then split at :229-231. Nothing seeds it from the agent.

The same key works one level up. packages/core/src/model-resolver.ts:154 merges
model.body with variant.body, and :253 merges credential metadata, so a temperature set on
a model or variant reaches the wire. The identical key on an agent does not, and there is no
warning either way.

Observed: an agent configured with temperature: 0.1 and maxOutputTokens: 32768 produced a
provider request carrying neither. Confirmed on the wire by the Magic Context plugin's test
harness against a mock provider; the same absence is visible in the source path above.

Expected: either agent-level request participates in request preparation the way model- and
variant-level overlays do, or — if it is deliberately inert on v2 — the field should be removed
from the agent schema so config that sets it fails loudly instead of doing nothing.

I have not filed a PR because the intended precedence between agent, model and variant overlays is
a design decision rather than an obvious fix, and I would rather have that answered first.