Fix relay-primary provider settlement durability

[YamonBot]

Provider deployment currently releases the managed-agent store lock while the external provider runs, then applies the returned handle to whichever relay-primary record is current. An A→Local or A→Provider-B edit during that interval can therefore attach A's handle to unrelated authority, and a retention failure can leave the only handle hidden behind an older encrypted head.

This change records an exact, scope-bound provider invocation before the external call and records any returned handle before settlement. Settlement requires the same private-head event, owner, relay scope, provider, and payload fingerprint; stale results remain visible as unresolved cleanup state and block ordinary redeploy/delete. The encrypted kind 30179 head is now committed before the device-local JSON mirror for provider results and ordinary agent edits.

It also includes the existing-cluster Secret reference needed by Kubernetes managed agents. The provider verifies the referenced Secret before mutation, loads it first through envFrom, and keeps Buzz's per-agent identity Secret authoritative.

Stacked on #4999 because it completes that PR's remaining provider settlement blockers.

Validation:

• just desktop-tauri-test — 3263 passed, 19 ignored; all workspace test targets passed
• cargo test -p buzz-backend-kubernetes — 160 unit and 4 wire tests passed
• just desktop-tauri-clippy
• CHECK_FILE_SIZES_BASE=6c1be331d86af8a02ad8e1454ff532e5a836a81a pnpm --dir desktop check:file-sizes
• git diff --check

[YamonBot]

Closing: the deployment must use an unmodified official upstream release and supported configuration path.