[proxy] enforce model allowlist for URL-routed providers (Bedrock/Vertex) (#6764)

## Describe your changes

The Agent Network policy Guardrail "Model Allowlist" was not enforced
for providers whose model travels in the URL/path rather than the JSON
body — most visibly AWS Bedrock (reported in netbirdio/netbird#6751),
and the same class applies to Google Vertex.

Root cause: the `llm_guardrail` allowlist check **failed open**.
`evaluateAllowlist` returned allow whenever the request model was absent
from the metadata bag (`middleware.go`, `if !modelPresent { return nil
}`). The model is stamped upstream by `llm_request_parser`; for
body-routed providers (OpenAI/Anthropic) it comes from the JSON body,
but for path-routed providers the model is recovered only when the
request matches a recognized path shape (Bedrock
`/model/{id}/{invoke|converse|...}`, Vertex
`/v1/projects/.../publishers/.../models/...`). Any shape the parser did
not recognize reached the guardrail with no model and was allowed
regardless of the allowlist.

Fix (provider-agnostic): **fail closed**. When an allowlist is
configured and the model cannot be determined (absent or empty), the
request is denied `403` with a distinct `llm_policy.model_unknown`
reason. This closes the bypass for Bedrock, Vertex, and any future
URL-routed provider in one place. When no allowlist is configured,
behavior is unchanged.

The model allowlist is enforced solely in the proxy `llm_guardrail`;
management's `CheckLLMPolicyLimits` handles only token/budget caps, so
no management change is required.

## Issue ticket number and link

<https://github.com/netbirdio/netbird/discussions/6751>

## Stack

- \#6726 <!-- branch-stack -->
  - \#6764 :point\_left:

### Checklist

- [x] Is it a bug fix
- [ ] Is a typo/documentation fix
- [ ] Is a feature enhancement
- [ ] It is a refactor
- [x] Created tests that fail without the change (if possible)
- [x] This change does **not** modify the public API, gRPC protocols,
functionality behavior, CLI / service flags, or introduce a new feature
— **OR** I have discussed it with the NetBird team beforehand (link the
issue / Slack thread in the description). See
[CONTRIBUTING.md](https://github.com/netbirdio/netbird/blob/main/CONTRIBUTING.md#discuss-changes-with-the-netbird-team-first).

> By submitting this pull request, you confirm that you have read and
agree to the terms of the [Contributor License
Agreement](https://github.com/netbirdio/netbird/blob/main/CONTRIBUTOR_LICENSE_AGREEMENT.md).

## Documentation

Select exactly one:

- [ ] I added/updated documentation for this change
- [x] Documentation is **not needed** for this change (explain why)

Bug fix that restores the documented allowlist behavior; no user-facing
surface changes.

### Docs PR URL (required if "docs added" is checked)

Paste the PR link from <https://github.com/netbirdio/docs> here:

<https://github.com/netbirdio/docs/pull/>\_\_

## Tests

- `llm_guardrail`: absent/empty model under a configured allowlist now
denies (`model_unknown`); empty allowlist still allows a missing model
(fail-closed only applies when a list is set); existing
allow/deny/case-insensitive cases retained.
- `llm_request_parser`: new parser→guardrail integration test drives
real **Bedrock** (`/model/{id}/invoke`) and **Vertex**
(`/v1/projects/.../models/...`) URL shapes and asserts allowed→200,
disallowed→403 (`model_blocked`), and an unrecognized Bedrock action→403
(`model_unknown`, the #6751 regression guard).

Note: a full through-tunnel e2e for the allowlist is intentionally
deferred — the agent-network e2e (`WaitProxyPeer`) is currently red on
`main`/`0.74.x` for an unrelated lazy-connection reason; it will be
added once that harness gate is fixed.

proxy/internal/middleware/builtin/llm_request_parser/guardrail_allowlist_test.go

@@ -0,0 +1,106 @@
+package llm_request_parser
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_guardrail"
+)
+
+// runParserGuardrail runs the request parser then the model-allowlist guardrail
+// in SlotOnRequest order, threading the parser's metadata into the guardrail the
+// same way the real chain does. It returns the guardrail decision so tests can
+// assert allowlist enforcement for URL/path-routed providers end to end.
+func runParserGuardrail(t *testing.T, url string, body []byte, allowlist []string) *middleware.Output {
+	t.Helper()
+	parser := newMiddleware(t)
+	parsed, err := parser.Invoke(context.Background(), &middleware.Input{
+		Slot: middleware.SlotOnRequest,
+		URL:  url,
+		Body: body,
+	})
+	require.NoError(t, err, "parser must not error")
+
+	guard := llm_guardrail.New(llm_guardrail.Config{ModelAllowlist: allowlist})
+	out, err := guard.Invoke(context.Background(), &middleware.Input{
+		Slot:     middleware.SlotOnRequest,
+		Metadata: parsed.Metadata,
+	})
+	require.NoError(t, err, "guardrail must not error")
+	require.NotNil(t, out, "guardrail must return an output")
+	return out
+}
+
+// TestModelAllowlist_URLRoutedProviders validates that the model allowlist is
+// enforced for providers whose model travels in the URL path (AWS Bedrock,
+// Google Vertex) rather than the JSON body. The "unknown action" case is the
+// regression guard for #6751: a Bedrock request shape the parser cannot map to a
+// model must fail closed under an allowlist instead of bypassing it.
+func TestModelAllowlist_URLRoutedProviders(t *testing.T) {
+	const bedrockBody = `{"anthropic_version":"bedrock-2023-05-31","messages":[{"role":"user","content":"hi"}]}`
+	const vertexBody = `{"anthropic_version":"vertex-2023-10-16","messages":[{"role":"user","content":"hi"}]}`
+
+	tests := []struct {
+		name      string
+		url       string
+		body      string
+		allowlist []string
+		decision  middleware.Decision
+		denyCode  string
+	}{
+		{
+			name:      "bedrock allowed model passes",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-haiku-4-5-v1:0/invoke",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionAllow,
+		},
+		{
+			name:      "bedrock disallowed model denied",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-opus-4-8-v1:0/invoke",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+		{
+			name:      "bedrock unknown action fails closed",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-opus-4-8-v1:0/some-future-action",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_unknown",
+		},
+		{
+			name:      "vertex disallowed model denied",
+			url:       "/v1/projects/p/locations/global/publishers/anthropic/models/claude-opus-4-8@20250101:rawPredict",
+			body:      vertexBody,
+			allowlist: []string{"claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+		{
+			name:      "vertex allowed model passes",
+			url:       "/v1/projects/p/locations/global/publishers/anthropic/models/claude-haiku-4-5@20250101:rawPredict",
+			body:      vertexBody,
+			allowlist: []string{"claude-haiku-4-5"},
+			decision:  middleware.DecisionAllow,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			out := runParserGuardrail(t, tt.url, []byte(tt.body), tt.allowlist)
+			assert.Equal(t, tt.decision, out.Decision, "unexpected decision for %s", tt.name)
+			if tt.decision == middleware.DecisionDeny {
+				require.NotNil(t, out.DenyReason, "deny reason must be set for %s", tt.name)
+				assert.Equal(t, 403, out.DenyStatus, "deny status must be 403 for %s", tt.name)
+				assert.Equal(t, tt.denyCode, out.DenyReason.Code, "deny code for %s", tt.name)
+			}
+		})
+	}
+}