feat(provider): support bounded auxiliary stream output ceilings

Problem

Server-owned auxiliary model calls, beginning with session-title generation, may need an enforceable output-token ceiling before a provider starts generation. The current title generator validates streamed usage after completion, so its 128-token threshold detects/rejects an oversized response but cannot prevent the provider from generating and billing those tokens.

port.LLMRequest must remain provider-neutral and must not grow provider-specific controls. Today the title generator receives an already constructed port.LLMProvider, so it has no composition/provider-construction seam through which to bind a provider-native output limit.

Requested capability

Design a reusable, composition-owned capability for bounded auxiliary streams that can apply a provider-native output limit at request construction without widening port.LLMRequest or exposing routing/credentials to callers. Preserve the stream-side usage check as defense in depth.

Initial consumer

internal/adapter/server/title_generator.go requests a bounded title response (128 output tokens) but currently checks the limit only after ChunkUsage.

Scope notes

• This is a correctness/cost-containment guard for pathological or misbehaving output, not a claim that normal title calls will exceed the limit.
• The solution should be reusable by future host-owned auxiliary calls.
• Keep provider selection, credentials, and provider-private configuration in composition/adapter construction; do not widen the core request DTO.