feat(orchestrator): per-Agent LLM model selection

[sneumannb5]

What

Per-Agent orchestrator LLM model selection (Closes #296). Each Agent can pin its own model (and per-turn triage routing) via the Admin builder instead of one
global ORCHESTRATOR_MODEL. Model picker is populated from the @omadia/llm-provider registry.

Why

Cost/latency differ per Agent — a high-volume support Agent can run on a cheap model while a reasoning-heavy one stays on the frontier model. Turns an
operator runtime choice into a per-instance setting instead of a deploy-time env decision. The pluggable-provider work only pays off once instances can
actually pick different models.

Test plan

• cd middleware && npm run typecheck -w @omadia/orchestrator
• cd middleware && npx tsc --noEmit in middleware and web-ui (clean)
• new/affected tests pass: configStoreModelRoutingValidation, agentBuilderAgentNode, agentGraphStoreSubAgentModelValidation, orchestratorRegistry, subAgentModelResolution, agentBuilderSubAgentTools
• manual: set per-Agent model in Admin builder → confirm effectiveModel badge + registry agent rebuilt log, turn runs without 404

Note: full middleware suite has 5 pre-existing failures in durableRecall/recallProbe — bisected to 1550268 (#385), date-relative test rot
(fixed 2026-06-01 data vs now()), unrelated to this PR.

Risk / blast radius

• No schema change — reuses existing agents.model_routing (JSONB) + agent_subagents.model.
• No new env var. ORCHESTRATOR_MODEL kept as the fallback tier → existing deployments unchanged.
• Resolution behaviour change (core): the orchestrator, per-turn router, and in-process sub-agents send model raw to a single provider adapter.
  Per-Agent refs are now resolved to the active provider's bare modelId before build (resolveModelIdForProvider): registry-known same-provider →
  modelId; cross-provider → dropped to default (out of scope, would 404 on the wrong adapter); registry-unknown → passed through raw (preserves
  pre-resolution behaviour, e.g. the default classifier).
• Public API (additive): GET /v1/builder/models now returns model_id (bare vendor id) alongside id.
• Hot-applies via the existing per-Agent rebuild — no restart.

Naming-decisions still pending

• Sub-agent model selection (SUB_AGENT_MODEL) is listed out-of-scope in Allow defining the LLM model per orchestrator instance #296 but this PR adds the sub-agent picker + validation. Confirm keep, or split out.