refactor(bedrock): replace apply_guardrail temp clone with a borrowed view

[vieiralucas]

Summary

`apply_guardrail` used to build a temporary `Guardrail` by field-cloning 13 values off either the DRAFT entry or a `GuardrailVersion`, only so that `evaluate_content` and the blocked-messaging fallback could read them back. Replace that with `GuardrailView<'a>`, a borrowed projection over the subset of fields the evaluator actually needs (`word_policy`, `topic_policy`, `sensitive_information_policy`, `blocked_input_messaging`, `blocked_outputs_messaging`).

• Add `GuardrailView::from_guardrail` and `GuardrailView::from_version` constructors.
• Extract a private `evaluate_content_view` that works over the view; the public `evaluate_content` still takes `&Guardrail` and delegates via `GuardrailView::from_guardrail` so its existing callers are unaffected.
• `apply_guardrail` now builds a view in each branch (DRAFT vs. versioned) and reads the blocked messaging straight from the view instead of a cloned temporary.

Other clone audit notes

• Kinesis `service.rs`: the `account_id = state.account_id.clone()` / `stream_name.clone()` pattern that shows up around 382 and 712 is load-bearing — the closure passing them into `stream_not_found` has to outlive the subsequent `state.streams.get_mut()` reborrow, so dropping the clones triggers the borrow checker. Left as-is.
• Cognito `service/misc.rs:648`, `service/users.rs:105`: the `pool_id`/`username`/`region`/`account_id` clones in the refresh-token and admin-create flows are similarly load-bearing — the owning values are read under one borrow and later consumed by token issuance + trigger delivery outside that borrow.

StepFunctions `interpreter.rs`, SecretsManager, and CloudFormation clone audits will be separate PRs.

Test plan

• `cargo fmt`
• `cargo clippy --workspace --all-targets -- -D warnings`
• `cargo test --workspace --exclude fakecloud-e2e --exclude fakecloud-conformance`

---

Summary by cubic

Replaced the temporary cloned Guardrail in apply_guardrail with a borrowed GuardrailView<'a> for content evaluation. This removes 13 field clones, reduces allocations, and keeps the public API and behavior unchanged.

• Refactors
  • Added GuardrailView with word_policy, topic_policy, sensitive_information_policy, blocked_input_messaging, blocked_outputs_messaging.
  • Added GuardrailView::from_guardrail and GuardrailView::from_version.
  • Introduced internal evaluate_content_view; evaluate_content(&Guardrail) now delegates to it.
  • Updated apply_guardrail to build a view (DRAFT or versioned) and read blocked messaging directly from it.

^{Written for commit 5634d5e. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 1 file