perf(kiro): cache provider-scoped model catalog to avoid per-account fetches

[tonybase]

Summary

When managing many Kiro accounts, model registration called ListAvailableModels once per account on every registration cycle — startup, config reload, and the 3-hour catalog refresh. With N accounts this amplified into N upstream requests per cycle (the kiro: fetched N models from API log line fires per account).

The Kiro model catalog is provider-scoped — identical for every account; profileArn only routes billing. So this PR caches the catalog once and shares it across all Kiro auths.

What changed

• Add a small TTL cache (kiroModelsCache) on Service for the Kiro base (pre-agentic) model list.
• fetchKiroModels now: cache lookup → single upstream ListAvailableModels (fetchKiroBaseModelsFromAPI) → per-read agentic variant generation.
• Coalescing is done by holding the mutex across the fetch: a concurrent burst of accounts collapses into one upstream call (latecomers block, then read the fresh cache). The TTL (30 min) collapses repeat fetches across registration cycles.
• Agentic variants are derived per read, not cached, because they depend on the kiro-system-prompt-inject-enable runtime flag rather than the account.
• Failed and empty fetches are not cached, so a later caller can retry (and still falls back to the static registry.GetKiroModels()).

Result

N Kiro accounts trigger at most one ListAvailableModels call per TTL, instead of one per account per registration cycle.

Notes / scope

• Scoped to Kiro only. The same per-account fetch amplification exists for cursor, github-copilot, kilo, qoder, and antigravity — if this approach is acceptable, it can be extended in a follow-up.
• The TTL is a constant (30 min). Happy to make it configurable if maintainers prefer.
• The shared cached slice is only consumed by non-mutating helpers (generateKiroAgenticVariants, applyExcludedModels, applyModelPrefixes all build new slices / clone before edits), so sharing is safe.

Testing

• New service_kiro_models_cache_test.go covering: concurrent coalescing (50 callers → 1 fetch), TTL reuse + expiry refresh, failures not cached, empty results not cached. All pass under -race.
• gofmt -w ., go build ./cmd/server, and go test ./sdk/cliproxy/... all pass.

🤖 Generated with Claude Code