[management,proxy] Agent network: per-account LLM gateway (policy, metering, multi-provider) (#6555)

* [agent-network] Shared proto, OpenAPI schema, and generated types

* [agent-network] Management: store, manager, synthesizer, policy engine, provider catalog, HTTP/gRPC API

Adds the account-scoped agent-network module: provider/policy/budget CRUD and
store, the reverse-proxy service synthesizer, policy selection + limit
enforcement, the provider catalog (incl. Vertex AI and AWS Bedrock entries),
and the management HTTP + proxy gRPC surfaces.

* [management] Fix agent-network proxy-peer fan-out on affected-peer recompute

The affected-peers resolver loaded only persisted reverse-proxy services, but
agent-network services are synthesized on demand and never persisted. As a
result the embedded proxy peer was never folded into the affected set when a
client's group changed, so the proxy received no network-map update for a newly
authorised client and rejected its handshake until a full resync (restart).

loadProxyServices now merges the synthesized agent-network services (injected
via a registration hook to avoid an import cycle), so proxy peers learn newly
authorised clients immediately.

* [proxy] Reverse-proxy middleware framework, chain, and request plumbing

The per-target middleware chain (slots, dispatcher, mutation gate, metadata
merger), body capture, access-log terminal sink, and the proxy wiring that
builds + runs chains for synthesized agent-network services.

* [proxy] LLM parsers, pricing, and builtin middlewares (OpenAI, Anthropic, Vertex AI, AWS Bedrock)

Request/response parsers and SSE/event-stream metering, the embedded pricing
table, and the builtin middleware set: request parser, router, policy
limit-check/record, cost meter, guardrail, identity inject, response parser.
Includes the path-routed providers — Google Vertex AI (keyfile:: service-account
OAuth minting) and AWS Bedrock (bearer auth, invoke/converse/streaming, optional
/bedrock prefix) — plus the Models allowlist and unmeterable-publisher deny.

* [proxy] IPv6 in-place apply and TCP accept-loop hardening on netstack listeners

* [agent-network] End-to-end test suite, module docs, and deployment preset

* [agent-network] Fix codespell typos and exclude false positives

- labelgen word pool: vermillion -> vermilion, racoon -> raccoon.
- codespell ignore list: add flate (Go compress/flate package), recordin
  (a test-local identifier), and unparseable (a valid alternative spelling used
  consistently across identifiers + a metadata-value constant).

* [management] Set LastSeen on injected proxy peer in realstack test (MySQL strict-mode)

The injected embedded proxy peer had a PeerStatus with a zero LastSeen, which
serializes to '0000-00-00' and is rejected by MySQL in strict mode (SQLite
tolerates it). Set LastSeen to a valid time so SaveAccount succeeds on both
engines.

* [agent-network] Remove e2e shell-script suite from this branch

The end-to-end shell scripts under scripts/e2e/ are maintained in a separate
testing suite and are not part of this change set.

* [agent-network] Polish module docs: remove internal review scaffolding, fix links, verify diagrams

Strip PR-review framing, commit references, absolute paths, and stale internal
references from the agent-network module docs; fix broken relative links; verify
all diagrams against the current architecture. Remove the internal AI-reviewer
prompt file.

* [management] Refine session expiration handling to support 3-state encoding for SSO deadlines

* [agent-network] Relocate agentnetwork package to internals/modules

Move management/server/agentnetwork (and its catalog/, labelgen/, types/
subpackages) to management/internals/modules/agentnetwork, alongside the
reverse-proxy module, and rewrite all importers. Pure relocation: package names,
the synthesizer + affectedpeers registration hook, and store access (shared
store.Store) are unchanged, so no import cycle is introduced (affectedpeers
still depends only on the agentnetwork/types leaf).

* [agent-network] Co-locate HTTP handlers in the module (RegisterEndpoints)

Move the agent-network HTTP handlers from server/http/handlers/agentnetwork into
the module at internals/modules/agentnetwork/handlers (package handlers) and
rename the entrypoint AddEndpoints -> RegisterEndpoints, matching the
reverse-proxy module convention. Wiring in http/handler.go updated accordingly.

proxy/internal/middleware/builtin/cost_meter/middleware.go

@@ -0,0 +1,193 @@
+// Package cost_meter implements the SlotOnResponse middleware that
+// converts token-usage metadata emitted by llm_response_parser into a
+// per-request USD cost estimate. The middleware uses the shared pricing
+// loader so operator pricing overrides apply to the chain.
+package cost_meter
+
+import (
+	"context"
+	"fmt"
+	"strconv"
+
+	"github.com/netbirdio/netbird/proxy/internal/llm/pricing"
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
+)
+
+// ID is the registry identifier for this middleware.
+const ID = "cost_meter"
+
+// Version is the implementation version emitted via the spec merge.
+const Version = "1.0.0"
+
+// Skip reasons emitted under KeyCostSkipped. The set is closed; the
+// dashboard surfaces these verbatim.
+const (
+	skipMissingProvider = "missing_provider"
+	skipMissingModel    = "missing_model"
+	skipMissingTokens   = "missing_tokens"
+	//nolint:gosec // skip-reason label, not a credential
+	skipUnparseableTokens = "unparseable_tokens"
+	skipZeroTokens        = "zero_tokens"
+	skipUnknownModel      = "unknown_model"
+)
+
+var metadataKeys = []string{
+	middleware.KeyCostUSDTotal,
+	middleware.KeyCostSkipped,
+}
+
+// Middleware computes a per-response cost estimate from the token
+// counts emitted upstream by llm_response_parser.
+type Middleware struct {
+	loader *pricing.Loader
+	// cancel stops this instance's pricing-reload goroutine. Non-nil only
+	// when the loader watches an override file; Close calls it so a chain
+	// rebuild doesn't leak a poll goroutine per retired instance.
+	cancel context.CancelFunc
+}
+
+// newMiddleware constructs a Middleware bound to the given pricing loader.
+// cancel may be nil (defaults-only loader with no reloader to stop).
+func newMiddleware(loader *pricing.Loader, cancel context.CancelFunc) *Middleware {
+	return &Middleware{loader: loader, cancel: cancel}
+}
+
+// ID returns the registry identifier.
+func (m *Middleware) ID() string { return ID }
+
+// Version returns the implementation version.
+func (m *Middleware) Version() string { return Version }
+
+// Slot reports that the middleware runs after the upstream call.
+func (m *Middleware) Slot() middleware.Slot { return middleware.SlotOnResponse }
+
+// AcceptedContentTypes is empty: cost_meter never inspects bodies.
+func (m *Middleware) AcceptedContentTypes() []string { return []string{} }
+
+// MetadataKeys returns the closed allowlist of keys this middleware
+// may emit.
+func (m *Middleware) MetadataKeys() []string {
+	return append([]string(nil), metadataKeys...)
+}
+
+// MutationsSupported reports that this middleware never mutates the
+// response.
+func (m *Middleware) MutationsSupported() bool { return false }
+
+// Close stops this instance's pricing-reload goroutine, if any. Called by
+// the chain when a rebuild retires the instance, so the mtime-poll loop
+// doesn't outlive the chain it belonged to. Safe to call on a nil receiver
+// and on an instance with no reloader.
+func (m *Middleware) Close() error {
+	if m != nil && m.cancel != nil {
+		m.cancel()
+	}
+	return nil
+}
+
+// Invoke reads provider, model, and token metadata, looks up pricing,
+// and emits either KeyCostUSDTotal or KeyCostSkipped. The decision is
+// always DecisionAllow; cost metering never denies or mutates.
+func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middleware.Output, error) {
+	out := &middleware.Output{Decision: middleware.DecisionAllow}
+	if in == nil {
+		return out, nil
+	}
+
+	provider := lookupKV(in.Metadata, middleware.KeyLLMProvider)
+	if provider == "" {
+		out.Metadata = skip(skipMissingProvider)
+		return out, nil
+	}
+
+	model := lookupKV(in.Metadata, middleware.KeyLLMModel)
+	if model == "" {
+		out.Metadata = skip(skipMissingModel)
+		return out, nil
+	}
+
+	inRaw, hasIn := lookupKVOK(in.Metadata, middleware.KeyLLMInputTokens)
+	outRaw, hasOut := lookupKVOK(in.Metadata, middleware.KeyLLMOutputTokens)
+	if !hasIn || !hasOut {
+		out.Metadata = skip(skipMissingTokens)
+		return out, nil
+	}
+
+	inTokens, err := strconv.ParseInt(inRaw, 10, 64)
+	if err != nil || inTokens < 0 {
+		// Unparseable or negative tokens are not a runtime error: the
+		// upstream llm_response_parser emitted a non-numeric / invalid
+		// value, so we surface that as cost.skipped and continue with
+		// Allow rather than pricing a negative count.
+		out.Metadata = skip(skipUnparseableTokens)
+		return out, nil //nolint:nilerr // structured skip; not a runtime error
+	}
+	outTokens, err := strconv.ParseInt(outRaw, 10, 64)
+	if err != nil || outTokens < 0 {
+		out.Metadata = skip(skipUnparseableTokens)
+		return out, nil //nolint:nilerr // structured skip; not a runtime error
+	}
+
+	// Cache buckets are optional and silently zeroed on a missing /
+	// malformed value; they're a refinement on top of input cost,
+	// not a precondition. A buggy value falls back to 0, never aborts.
+	cachedTokens := parseOptionalInt64(in.Metadata, middleware.KeyLLMCachedInputTokens)
+	cacheCreationTokens := parseOptionalInt64(in.Metadata, middleware.KeyLLMCacheCreationTokens)
+
+	if inTokens == 0 && outTokens == 0 && cachedTokens == 0 && cacheCreationTokens == 0 {
+		out.Metadata = skip(skipZeroTokens)
+		return out, nil
+	}
+
+	table := m.loader.Get()
+	cost, ok := table.Cost(provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
+	if !ok {
+		out.Metadata = skip(skipUnknownModel)
+		return out, nil
+	}
+
+	out.Metadata = []middleware.KV{
+		{Key: middleware.KeyCostUSDTotal, Value: fmt.Sprintf("%.6f", cost)},
+	}
+	return out, nil
+}
+
+// skip returns a single-entry metadata slice carrying the given skip
+// reason under KeyCostSkipped.
+func skip(reason string) []middleware.KV {
+	return []middleware.KV{{Key: middleware.KeyCostSkipped, Value: reason}}
+}
+
+// lookupKV returns the value associated with key, or the empty string
+// when the key is absent.
+func lookupKV(kvs []middleware.KV, key string) string {
+	v, _ := lookupKVOK(kvs, key)
+	return v
+}
+
+// lookupKVOK returns the value associated with key plus a presence
+// flag so callers can distinguish absent from empty.
+func lookupKVOK(kvs []middleware.KV, key string) (string, bool) {
+	for _, kv := range kvs {
+		if kv.Key == key {
+			return kv.Value, true
+		}
+	}
+	return "", false
+}
+
+// parseOptionalInt64 reads a metadata value and decodes it as int64.
+// Absent or unparseable values yield 0 — the caller treats absence as
+// "no cached tokens" rather than an error, since cache buckets are a
+// refinement, not a precondition.
+func parseOptionalInt64(kvs []middleware.KV, key string) int64 {
+	raw, ok := lookupKVOK(kvs, key)
+	if !ok {
+		return 0
+	}
+	v, err := strconv.ParseInt(raw, 10, 64)
+	if err != nil || v < 0 {
+		return 0
+	}
+	return v
+}