#2172)

Follow-up to #2183 per multi-model design principle: LlmPlanner now
accepts a dedicated provider reference instead of always using the
primary provider.

Config field `planner_provider = "fast"` in `[orchestration]`
references a `[[llm.providers]]` entry by name; empty string falls
back to the primary provider. Field replaces the dead
`planner_model: Option<String>` (never wired, pre-v1.0.0 removal).

`build_planner_provider()` in bootstrap resolves the named provider
and passes it to `LlmPlanner::new()` at all three wiring sites
(daemon, runner, acp). `OrchestrationState.planner_provider` holds
the optional resolved provider, scoped to orchestration rather than
the global ProviderState.

`migrate_planner_model_to_provider()` auto-migration step comments
out any existing `planner_model` value with a MIGRATED marker so
users know to supply a provider name instead of a model name.

Also adds latency-threshold doc comment on
`TopologyClassifier::suggest_max_parallel` (parallel scheduling
overhead only pays off when individual tool calls take >= 500ms;
for LLM API calls this is always satisfied; for local tools such as
file reads, consider keeping topology_selection disabled).

+50 tests (6504 -> 6554).