fix(models): provider-scoped allow-list + real capabilities for relay-fronted models

[spytensor]

Two related fixes for running MOZI behind a relay (LiteLLM / OpenRouter / corporate gateway).

1. Cross-provider allow-list collision (data bug)

A relay re-lists another provider's model id — e.g. deepseek-v4-flash under both DeepSeek and Qwen. The allow-list was keyed by bare model id (tenant_quotas.allowed_models, users.allowed_models, the UI activation Set, and every runtime match site), so removing the model from one provider silently removed it from the other. Reported first-hand.

Fix: key the allow-list by `${provider}:${model}` (provider ids never contain :, so the first colon splits unambiguously even for bedrock-style ...v2:0 ids).

• entitlements: makeModelEntitlementKey + composite/legacy-bare matchers; isModelAllowed/assertModelAllowed take providerId.
• model-router passes selection.provider; provider-failover filters via entry.provider.
• application-routes: buildProviderModels + validateKnownModelGrant understand composite keys.
• UI: activation state keyed by modelActivationKey(provider, model) throughout.

No DB/DDL migration — the column stays a JSON string[]; format is extended and reads stay backward-compatible with legacy bare-id lists (a bare entry matches any provider until the operator next saves, which rewrites in composite form).

2. Relay-fronted models lost their real capabilities

resolveRuntimeModel returned a blanket conservative profile (context 32k, max output 4k, no tools, no vision) for any id not in the built-in catalog — so a relay-fronted 200k-context vision model was silently capped and had vision routing disabled.

MOZI already downloads LiteLLM's own price/context registry (model-registry-enrichment) but only used it in the API display layer, not in the live resolver that feeds the brain's context sizing (handler) and vision routing (model-router).

Fix: wire getCachedModelMetadata into resolveRuntimeModel's unknown-model branch, so every consumer that resolves through it gets the model's real context window, tool/vision support and pricing. Conservative profile still applies when the registry has no entry (a private id) or the cache is cold. No schema, UI, or migration.

Tests

• entitlement-key.test.ts — composite matcher isolation + legacy compat + colon-in-id. ✅
• SettingsView.test.tsx — two-provider same-id regression: removing DeepSeek's copy keeps Qwen's; saved list is ["dashscope:deepseek-v4-flash"]. ✅ 24/24
• relay-model-enrichment.test.ts — relay-fronted id resolves to real caps from a seeded registry; conservative fallback intact. ✅
• Typecheck clean on backend + UI; touched provider/model tests green (54).

Not verified here

Backend DB-integration tests could not run locally: this machine's Node is v26, outside the repo's <26 engines range, and better-sqlite3 fails to build against Node 26's V8. The DB-resolution path is wiring-unverified at runtime on this machine — confirm in CI / a supported Node.