feat(platform): model router core — class config (fast|strong) over the LLM seam #105

Goal: a thin TS router (platform/model-router) where callers request a
class (fast | strong), not a vendor. Config maps class → { model, fallback[], max_budget }, resolved through the existing LLM factory
(apps/api/src/modules/memory/llm/index.ts). Per ADR-0002.

How to do it

• Define ModelClass = 'fast' | 'strong' and a config map → provider+model+
  fallback chain (fast → gpt-4o-mini/haiku; strong → sonnet, fallback opus).
• router.complete({ class, ... }) resolves the class and delegates to the
  current provider seam; with no API key every class resolves to the fake
  provider (CI stays offline).
• Hot-swap model via config, no call-site change.

Acceptance criteria

• router.complete({ class: 'fast' }) resolves to the cheap model; 'strong'
  to the strong model; both fall back to fake with no key.
• Hermetic test asserts class→model resolution via the fake provider.

🔗 Sequencing

• ⛔ Blocked by: arch(platform): model router + cost chokepoint (LiteLLM-style class routing) (§09/§12) #75, roadmap: Phase 0 exit-gate — the dogfood criteria that unlock Phase 1 #73 — do these first.
• 🚧 Blocks: feat(platform): per-user + per-route budget caps (reject over-budget calls) #106, feat(platform): per-call cost computation + telemetry in /admin/stats #107, feat(platform): wire encoder→fast, recall→strong + assert ingest is fast-only #108 — these wait on this one.