feat(up): Foundry auto-setup, best-model selection, memory CRD parity + fix kars up hang

[pallakatos]

Summary

Makes kars up --foundry-endpoint actually set up a BYO Foundry project so Memory Store works out of the box, stops hardcoding a stale model, fixes the post-deploy hang, and surfaces previously-masked memory errors.

Surfaced from real clean-room runs (Pal + @laevenso). Not for merge until tested.

Foundry auto-setup (new cli/src/commands/up/foundry_setup.ts)

• Discover the project; list deployed models via ARM control-plane (caller's own az token — no Graph).
• Best deployed chat model instead of hardcoded gpt-4.1 (pure, tested ranking; --model always wins; excludes embedding/image/audio). On the live azureclaw-foundry set this picks gpt-5.4.
• Ensure an embedding model (Memory Store needs one) — best-effort deploy text-embedding-3-small if absent.
• Enable the project system-assigned MI if missing (Memory Store authenticates internally as the project MI), then re-read its principalId for the existing Azure AI User RBAC grant.
• All idempotent + non-fatal — every failure degrades to a clear note; the deploy never aborts.

CRD parity + status

• Emit a KarsMemory binding CR on kars up (Foundry endpoints only), matching what kars dev already creates.
• CRD status report after apply (InferencePolicy / ToolPolicy / KarsMemory / KarsSandbox + phase).

Fix the kars up hang (two causes)

• cli/src/preflight.ts: the RBAC spinner was only concluded when fetchSubscriptionPermissions threw or returned a non-empty set. An empty [] (no throw) left it spinning — its setInterval kept Node alive, so kars up hung after the summary with the spinner still animating (reproduced by two operators). Now concluded on the empty path. A second identical leak in the provider notFound path is also fixed.
• up.ts: process.exit(0) on success (belt-and-suspenders for the detached kubectl port-forward handle).

Memory error unmasking (runtime)

• ensureStore uses the strict router call for POST /memory_stores so the real 403/400 surfaces (MI not enabled / RBAC still propagating / no embedding model) instead of the generic "could not be created".

Security audit

docs/internal/security-audits/2026-06-25-foundry-autosetup-bestmodel-memory-spinner.md (2 sign-offs). No new role/scope/principal — the two writes are operator-scoped, idempotent, best-effort, on their own Foundry resource. security-audit-required + copyright-headers pass locally.

Verification

• CLI: tsc clean, oxlint 0 errors, 831 tests (+10 new).
• Runtime: tsc clean, oxlint 0 errors, 244 tests.

Note

The memory-unmask change lives in the sandbox image — needs kars push --only sandbox --apply (or the release build) to reach a running pod. The kars up changes are CLI-only and effective immediately.

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None