@namzu/sandbox 14.0.0 → 16.0.0

package/src/backends/kubernetes/index.ts

@@ -77,16 +77,35 @@
  * ## Egress
  *
  * `config.egress` is optional and, when set, translated and VERIFIED — never
- * created — by `egress-policy.ts`. Verification happens once, lazily, on the
- * first `create()`, so `buildKubernetesBackend` itself still contacts
- * nothing. Every Sandbox this file creates directly (`buildSandboxBody`)
- * carries {@link sandboxTemplateLabel} on its podTemplate specifically so
- * that translated policy's `podSelector` has something stable to match —
- * see `objects.ts`'s doc comment on that label for why agent-sandbox's own
- * controller-owned label does not cover this path.
+ * created — by `egress-policy.ts`, in two steps. The NAMED object is GETted
+ * and compared to the translation exactly, once, lazily, on the first
+ * `create()`, so `buildKubernetesBackend` itself still contacts nothing.
+ * Then, because the API server UNIONS every policy selecting a pod, every
+ * `NetworkPolicy` in the namespace (and, under `engine: 'cilium'`, every
+ * `CiliumNetworkPolicy`) is enumerated against the pod's real labels and the
+ * create is refused when any of them lets out more than the translation does
+ * — a second policy widens egress however exactly the named one matches, and
+ * a `SandboxTemplate`'s own `networkPolicy` block becomes exactly such a
+ * policy. `egress.verify: 'named-object-only'` is the opt-out and restores
+ * the first step alone. Every Sandbox this file creates directly
+ * (`buildSandboxBody`) carries {@link sandboxTemplateLabel} on its
+ * podTemplate specifically so that translated policy's `podSelector` has
+ * something stable to match — see `objects.ts`'s doc comment on that label
+ * for why agent-sandbox's own controller-owned label does not cover this
+ * path.
+ *
+ * ## Ingress
+ *
+ * `config.ingress` is the opposite default: verification is ON unless a
+ * deployment says `'unverified'`. Before the POST for a direct Sandbox, and
+ * after the bind for a claimed one, `ingress-policy.ts` lists the namespace's
+ * policies and refuses unless one of them actually closes the agent port on
+ * the labels this pod carries. Nothing checked that before, while three
+ * pieces of shipped text said it was covered — see that module's own doc
+ * comment for what was measured.
  */
 
-import type { Sandbox } from '@namzu/sdk'
+import type { Sandbox, SandboxNetworkPolicy } from '@namzu/sdk'
 import { generateSandboxId } from '@namzu/sdk'
 
 import type { SandboxBackend, SandboxBackendOptions } from '../../index.js'
@@ -97,46 +116,88 @@ import {
 	runFailureCleanup,
 } from '../readiness.js'
 import {
+	type EgressProfileLabel,
 	type KubernetesEgressConfig,
+	KubernetesPodLabelNotObservedError,
+	KubernetesPodLabelsRejectedError,
+	type KubernetesTranslatedEgressPolicy,
 	assertEgressPolicyIsEnforceable,
+	assertEgressProfileIsUsable,
+	assertPerSandboxEgressIsUsable,
+	composeAdditionalPodLabels,
 	defaultEgressPolicyName,
+	egressProfileLabel,
+	egressUnionVerificationEnabled,
+	perSandboxEgressLabelKey,
 	translateEgressPolicy,
 	verifyEgressPolicyApplied,
+	verifyEgressPolicyUnion,
 } from './egress-policy.js'
+import {
+	type KubernetesIngressConfig,
+	ingressVerificationEnabled,
+	resolveIngressEngine,
+	verifyIngressPolicyApplied,
+} from './ingress-policy.js'
 import {
 	type KubernetesAccess,
 	KubernetesAlreadyGoneError,
+	KubernetesApiError,
+	KubernetesApiTimeoutError,
 	type KubernetesClient,
+	type KubernetesClientOptions,
+	KubernetesCredentialError,
 	createKubernetesClient,
 } from './k8s-client.js'
 import {
+	type KubernetesCondition,
 	type PodListResource,
 	type PodResource,
 	READY_CONDITION,
 	SANDBOX_API_GROUP,
 	SANDBOX_API_VERSION,
 	SANDBOX_EXTENSIONS_API_GROUP,
+	type SandboxClaimListResource,
 	type SandboxClaimResource,
 	type SandboxPodTemplate,
 	type SandboxResource,
 	type SandboxTemplateResource,
 	type SandboxVolumeClaimTemplate,
+	type SandboxWarmPoolResource,
 	claimCollectionPath,
+	claimListPath,
 	claimPath,
 	isConditionTrue,
 	isPodLive,
+	podCollectionPath,
 	podListPath,
 	podPath,
+	readPodIP,
 	sandboxCollectionPath,
 	sandboxPath,
 	sandboxTemplateLabel,
 	sandboxTemplatePath,
+	warmPoolPath,
 } from './objects.js'
+import {
+	KubernetesOwnerUidMissingError,
+	type PerSandboxPolicyOwner,
+	buildAdmissionFence,
+	buildPerSandboxPolicySetter,
+} from './per-sandbox-policy.js'
 import { privilegeProbeTimedOut, runPrivilegeProbe } from './privilege-probe.js'
 import { buildKubernetesSandbox } from './sandbox.js'
 import { KubernetesAgentTransport } from './transport.js'
 
-export type { KubernetesEgressConfig, KubernetesEgressEngine } from './egress-policy.js'
+export type {
+	KubernetesEgressConfig,
+	KubernetesEgressEngine,
+	KubernetesEgressPolicy,
+	KubernetesEgressVerification,
+	KubernetesOnlyEgressPolicy,
+	KubernetesPerSandboxEgressConfig,
+} from './egress-policy.js'
+export type { KubernetesIngressConfig, KubernetesIngressEngine } from './ingress-policy.js'
 
 /**
  * How the backend reaches the API server. Two sources, neither needing a YAML
@@ -166,6 +227,13 @@ export interface KubernetesBackendInternalConfig {
 	readonly warmPoolName?: string
 	/** TCP port the guest agent listens on. Default {@link DEFAULT_AGENT_PORT}. */
 	readonly agentPort?: number
+	/**
+	 * Which of a sandbox's two addresses the transport dials. Default
+	 * `'service'` — see {@link KubernetesAgentAddressMode}, which is where
+	 * the choice is explained, because it is a fact about where the HOST
+	 * runs rather than about the cluster.
+	 */
+	readonly agentAddress?: KubernetesAgentAddressMode
 	readonly readyPollIntervalMs?: number
 	readonly readyTimeoutMs?: number
 	/**
@@ -189,13 +257,124 @@ export interface KubernetesBackendInternalConfig {
 	/**
 	 * Egress policy this backend's `NetworkPolicy` (or `CiliumNetworkPolicy`,
 	 * under `engine: 'cilium'`) is expected to carry. Unset means this backend
-	 * neither computes nor verifies one — the cluster's default posture (the
-	 * SandboxTemplate's own managed NetworkPolicy) is all that applies. See
-	 * `egress-policy.ts`.
+	 * neither computes nor verifies one, and OUTBOUND traffic is then whatever
+	 * the cluster's own policies happen to allow.
+	 *
+	 * Deliberately not described as "covered by the SandboxTemplate's managed
+	 * NetworkPolicy", which is what this comment used to claim: that policy
+	 * selects `agents.x-k8s.io/sandbox-template-ref-hash`, a label the
+	 * controller writes only onto a Sandbox adopted out of a `SandboxWarmPool`
+	 * and never onto one this backend POSTs — so for every pool-less sandbox
+	 * and every workspace it selects nothing at all. See `ingress-policy.ts`.
 	 */
 	readonly egress?: KubernetesEgressConfig
+	/**
+	 * Whether the agent port's INGRESS boundary is verified before a sandbox
+	 * is created, and against which policy resources.
+	 *
+	 * Unset means VERIFY — the one field in this config whose absent value is
+	 * the strict one, because the deployment that needs the check is the one
+	 * that would never have set it. `'unverified'` reads no policy and issues
+	 * no request, for a deployment whose boundary lives somewhere a namespaced
+	 * Role cannot see. See `ingress-policy.ts`.
+	 */
+	readonly ingress?: KubernetesIngressConfig
+	/**
+	 * Bound on every single Kubernetes API request this backend sends. See
+	 * {@link KubernetesClientOptions.requestTimeoutMs} in `k8s-client.ts`,
+	 * which owns the default (30 s), the floor (1 s) and the reason there is
+	 * no value that disables it.
+	 */
+	readonly apiRequestTimeoutMs?: number
+	/**
+	 * Interval of the negotiated liveness heartbeat on `openTerminal` and
+	 * `openTcpConnection` streams. Default
+	 * {@link DEFAULT_STREAM_HEARTBEAT_MS}; `0` turns it off and restores the
+	 * pre-heartbeat behaviour exactly. See {@link resolveStreamHeartbeatMs}.
+	 */
+	readonly streamHeartbeatMs?: number
+	/**
+	 * Extra labels written onto every `SandboxClaim` this backend POSTs —
+	 * `metadata.labels`, and nowhere else. Never merged into
+	 * `additionalPodMetadata`: those are POD labels a running Sandbox and its
+	 * `NetworkPolicy` selectors read, and a host's own recovery bookkeeping
+	 * has no business changing what a pod is selected by. Absent means no
+	 * labels beyond what the controller itself writes, and every claim body
+	 * this backend sends is byte-for-byte what it always was.
+	 *
+	 * The intended use is a host-instance identity — e.g.
+	 * `{ 'sandbox.namzu.ai/host-instance': hostId }` — so a restarted host
+	 * can find and {@link releaseKubernetesTaskSandboxes} its predecessor's
+	 * claims well before `claimTtlSeconds` would reap them on its own.
+	 */
+	readonly claimLabels?: Record<string, string>
+}
+
+/**
+ * Default {@link KubernetesBackendInternalConfig.streamHeartbeatMs} — 15 s,
+ * so a stream whose peer vanished without a FIN or an RST is given up on
+ * within 45 s rather than never.
+ *
+ * The Kubernetes backend opts IN here; `VsockTransportOptions.heartbeatMs`
+ * stays undefined by default, because that transport is shared with the
+ * Firecracker tier and a default there would force-close an existing
+ * consumer's quiet-but-alive terminal.
+ */
+export const DEFAULT_STREAM_HEARTBEAT_MS = 15_000
+
+/**
+ * Validate the configured stream-heartbeat interval, or supply the default.
+ *
+ * `0` IS accepted here, unlike `apiRequestTimeoutMs`: the heartbeat is a new
+ * capability that a deployment behind a middlebox with its own idea about
+ * unexpected frames may want off, and turning it off restores exactly the
+ * behaviour every release before this one had. An unanswered API request has
+ * no such prior behaviour worth restoring.
+ */
+export function resolveStreamHeartbeatMs(value: number | undefined): number {
+	if (value === undefined) return DEFAULT_STREAM_HEARTBEAT_MS
+	if (!Number.isSafeInteger(value) || value < 0) {
+		throw new Error(
+			`kubernetes: streamHeartbeatMs must be a non-negative integer, got ${JSON.stringify(
+				value,
+			)}. Use 0 to send no heartbeats at all, which is how every release before this one behaved; the default is ${DEFAULT_STREAM_HEARTBEAT_MS}ms.`,
+		)
+	}
+	return value
 }
 
+/**
+ * The client options every `createKubernetesClient` call in this backend is
+ * built with. One function so the five call sites — the provider, and each
+ * of the workspace verbs — cannot drift apart on which bounds they honour.
+ */
+export function clientOptions(config: KubernetesBackendInternalConfig): KubernetesClientOptions {
+	return { requestTimeoutMs: config.apiRequestTimeoutMs }
+}
+
+/**
+ * Which address a sandbox's agent is dialed at. A property of where the HOST
+ * runs, not of the cluster.
+ *
+ *  - `'service'` (default) — the Sandbox's `status.serviceFQDN`,
+ *    `<name>.<namespace>.svc.cluster.local`. It outlives the pod: a resumed
+ *    workspace comes back behind the same name, and every dial re-resolves
+ *    it. The catch is that only the cluster's own DNS answers it, so this is
+ *    correct exactly when the host itself runs inside the cluster.
+ *  - `'pod-ip'` — the bound pod's IP, read from the same `GET` that reads
+ *    its uid, so the address and the bind token are always one pod's. For a
+ *    host OUTSIDE the cluster on a routable pod network (a peered VNet, a
+ *    node-local operator, a CI runner with a route): it needs no cluster
+ *    resolver at all. The cost is that a pod IP dies with its pod, which is
+ *    why a resume re-reads it and a connect failure re-reads it once.
+ *
+ * The cluster still decides whether either address is REACHABLE: `'pod-ip'`
+ * additionally needs a NetworkPolicy that admits the host's own address
+ * range on the agent port. Neither mode changes the bind token, the
+ * privilege probe or egress verification.
+ */
+export type KubernetesAgentAddressMode = 'service' | 'pod-ip'
+
 /**
  * The address the guest agent answers on, plus the token to present.
  *
@@ -224,8 +403,33 @@ export interface KubernetesSandboxBinding {
 export interface KubernetesAcquisition {
 	readonly binding: KubernetesSandboxBinding
 	readonly agent: KubernetesAgentAddress
+	/**
+	 * Re-read the live pod and recompute {@link agent} from it. Present only
+	 * under `agentAddress: 'pod-ip'`, where the address is a literal that
+	 * dies with its pod; a Service FQDN needs no such thing, and handing one
+	 * over anyway would give the default mode a re-read it never had.
+	 */
+	readonly refreshAgent?: (signal?: AbortSignal) => Promise<KubernetesAgentAddress>
 	/** API path of the object THIS backend created — the claim, or the Sandbox. */
 	readonly ownedPath: string
+	/**
+	 * The object this backend created, identified well enough to be named in
+	 * another object's `ownerReferences`: its kind, its name and the uid the
+	 * API server assigned it.
+	 *
+	 * Present only when `config.egress.perSandbox` is configured, because it
+	 * is read from the create reply and the readiness polls and nothing else
+	 * needs it — a deployment that never writes a per-sandbox policy should
+	 * not start carrying a field whose absence would otherwise be a bug.
+	 */
+	readonly owner?: PerSandboxPolicyOwner
+	/**
+	 * The VALUE of the per-sandbox selector label on this sandbox's pod —
+	 * the created object's name, confirmed on the bound pod before the
+	 * sandbox was admitted. Present under the same condition as
+	 * {@link owner}.
+	 */
+	readonly perSandboxLabelValue?: string
 	/** The TTL acquire stamped, which every renewal re-stamps. */
 	readonly ttlSeconds: number
 	/** DELETE that object. An already-gone object counts as released. */
@@ -378,69 +582,311 @@ export function buildKubernetesBackend(config: KubernetesBackendInternalConfig):
 	// `config.egress.policy.kind` alone, with no API call, so it is refused
 	// here, synchronously, the same moment the two checks above are.
 	if (config.egress) {
-		assertEgressPolicyIsEnforceable(config.egress.policy, config.egress.engine ?? 'core')
+		assertEgressPolicyIsEnforceable(
+			config.egress.policy,
+			config.egress.engine ?? 'core',
+			config.egress.ciliumNarrowing,
+		)
 	}
-	const client = createKubernetesClient(clientAccess(config))
+	// And the same for an egress PROFILE that could never be written as a
+	// label — a value the API server would reject leaves either a claim
+	// nothing binds or a policy nobody can apply, and both are decidable
+	// from config alone.
+	assertEgressProfileIsUsable(config.egress)
+	// And the same for per-sandbox egress: an engine that cannot express a
+	// hostname, an unnamed admission fence or a selector key the API server
+	// would refuse are all decidable from config alone, and a host that
+	// learns any of them from its first `setNetworkPolicy` call learns it an
+	// hour into a run that cannot be redone.
+	assertPerSandboxEgressIsUsable(config.egress)
+	// Resolved here as well as at each session, so a configuration this
+	// backend will never honour is refused while `buildKubernetesBackend` is
+	// still on the stack rather than on someone's first `create()`.
+	resolveStreamHeartbeatMs(config.streamHeartbeatMs)
+	const client = createKubernetesClient(clientAccess(config), clientOptions(config))
 	// Verify-not-trust runs once, lazily, on the first `create()` — never here,
 	// because `buildKubernetesBackend` is documented to contact nothing. A
 	// failed attempt is not cached: a transient API error should not wedge
-	// every later create() behind the same stale rejection forever.
-	let egressVerification: Promise<void> | undefined
+	// every later create() behind the same stale rejection forever. The
+	// boundary object holds both halves and both memos; it lives as long as
+	// this backend does, which is what makes the named-object check
+	// once-per-backend rather than once-per-create.
+	const egressBoundary = buildEgressBoundary(client, config, config.sandboxTemplateName)
+	// Ingress is cached PER LABEL SET rather than once per backend, because
+	// unlike the egress NAMED-object check it is a question about one pod: a
+	// pooled sandbox's labels come off the pool's template and a pool-less
+	// one's off this config, and a single memo would answer for a pod it never
+	// examined. Same failure handling as the egress memo — a failed attempt is
+	// dropped, so a transient API error does not wedge every later create()
+	// behind it. The egress UNION check is keyed the same way, for the same
+	// reason, and additionally expires: see {@link EGRESS_UNION_CACHE_TTL_MS}.
+	const ingressVerifier = buildIngressVerifier(client, config, new Map())
+	// One fence per backend, so its memo is shared by every sandbox this
+	// backend hands out rather than re-proved per handle. `undefined` when
+	// per-sandbox egress is not configured, which is what makes
+	// `setNetworkPolicy` absent from the handle — presence follows
+	// CONFIGURATION and never a runtime probe, so a caller's capability
+	// detection cannot depend on when it asked.
+	const perSandbox = config.egress?.perSandbox
+	const fence = perSandbox === undefined ? undefined : buildAdmissionFence(client, perSandbox)
 	return {
 		tier: 'microvm',
 		name: 'kubernetes',
 		async create(options: SandboxBackendOptions): Promise<Sandbox> {
-			if (config.egress) {
-				egressVerification ??= verifyEgressPolicyConfigured(
-					client,
-					config.namespace,
-					config.sandboxTemplateName,
-					config.egress,
-					options.signal,
-				).catch((err: unknown) => {
-					egressVerification = undefined
-					throw err
-				})
-				await egressVerification
-			}
-			const acquisition = await acquireKubernetesSandbox(client, config, options, readiness)
+			await egressBoundary?.verifyNamedObject(options.signal)
+			const acquisition = await acquireKubernetesSandbox(
+				client,
+				config,
+				options,
+				readiness,
+				ingressVerifier,
+				egressBoundary,
+			)
+			const egress = config.egress
+			const setNetworkPolicy =
+				fence !== undefined &&
+				egress?.perSandbox !== undefined &&
+				acquisition.owner !== undefined &&
+				acquisition.perSandboxLabelValue !== undefined
+					? buildPerSandboxPolicySetter({
+							client,
+							fence,
+							namespace: config.namespace,
+							egress: { ...egress, perSandbox: egress.perSandbox },
+							owner: acquisition.owner,
+							selectorValue: acquisition.perSandboxLabelValue,
+						})
+					: undefined
 			return await admitProbedSandbox(
 				acquisition,
 				config,
 				options,
 				resolveProbeTimeoutMs(readiness.timeoutMs),
+				setNetworkPolicy,
 			)
 		},
 	}
 }
 
 /**
- * Translate `egress.policy` and confirm an operator applied a matching
- * object — the whole verify-not-trust step, isolated so `create()` above
- * stays about ONE thing (memoize-once-per-backend) rather than two.
+ * The two egress checks a create path runs, with their memos.
+ *
+ * `undefined` when `config.egress` is unset — the whole boundary is one
+ * absent object rather than a flag every call site re-reads, the same shape
+ * {@link IngressVerifier} uses for its own opt-out.
  *
- * Exported because `workspace.ts` runs the identical step: a workspace does
+ * Exported because `workspace.ts` runs the identical steps: a workspace does
  * not go through `buildKubernetesBackend`, and a config `egress` honoured on
  * one entry point and ignored on the other would be a silent downgrade of the
- * boundary this backend calls primary. `sandboxTemplateName` is the template
- * the caller is actually building from — it decides both the default policy
- * name and the pod label the policy's selector has to match, and a workspace
- * may be built from a different template than the task path's.
+ * boundary this backend calls primary. It builds its own boundary per create,
+ * which is what makes its checks per-call rather than memoized — creating a
+ * workspace is a rare, explicit act with nothing to amortise, and a policy
+ * deleted since the last call must be noticed.
  */
-export async function verifyEgressPolicyConfigured(
+export interface KubernetesEgressBoundary {
+	/**
+	 * Translate `egress.policy` and confirm an operator applied a matching
+	 * object — the original verify-not-trust step, unchanged, including its
+	 * exact-match comparison and its once-per-boundary memo.
+	 */
+	verifyNamedObject(signal?: AbortSignal): Promise<void>
+	/**
+	 * Enumerate every policy selecting THIS pod and refuse when any of them
+	 * allows egress the translation does not. A no-op under
+	 * `egress.verify: 'named-object-only'`.
+	 */
+	verifyUnion(
+		podLabels: Readonly<Record<string, string>>,
+		subject: string,
+		signal?: AbortSignal,
+	): Promise<void>
+}
+
+/**
+ * How long a union pass is trusted for one label set.
+ *
+ * Five minutes rather than the backend's lifetime, which is what the
+ * named-object check alone used to get: an operator who applies a widening
+ * policy at 10:00 should not have it go unnoticed until the host restarts.
+ * It is a cache, not a watch — `k8s-client.ts`'s "no watch, no informers, no
+ * resourceVersion tracking" invariant is untouched, because the only thing
+ * kept across calls is "this exact label set passed at this time".
+ */
+export const EGRESS_UNION_CACHE_TTL_MS = 5 * 60 * 1_000
+
+/** One cached pass, and when it was taken. */
+interface CachedPass {
+	readonly at: number
+	readonly pending: Promise<void>
+}
+
+/**
+ * Build the egress boundary this config asks for, or nothing at all.
+ *
+ * `sandboxTemplateName` is the template the caller is actually building from
+ * — it decides both the default policy name and the pod label the policy's
+ * selector has to match, and a workspace may be built from a different
+ * template than the task path's.
+ *
+ * `now` is injected only so the TTL above can be tested without waiting five
+ * minutes; nothing else passes it.
+ */
+export function buildEgressBoundary(
 	client: KubernetesClient,
-	namespace: string,
+	config: KubernetesBackendInternalConfig,
 	sandboxTemplateName: string,
-	egress: KubernetesEgressConfig,
-	signal?: AbortSignal,
-): Promise<void> {
+	now: () => number = Date.now,
+): KubernetesEgressBoundary | undefined {
+	const egress = config.egress
+	if (egress === undefined) return undefined
 	const engine = egress.engine ?? 'core'
-	const translated = await translateEgressPolicy(egress.policy, engine, {
-		namespace,
-		name: egress.networkPolicyName ?? defaultEgressPolicyName(sandboxTemplateName),
+	// One resolution of the profile, shared by the policy NAME and the policy
+	// SELECTOR: under a profile the default name gains the profile segment
+	// (one template under two profiles is two policy objects) and the
+	// selector gains the label, and the two must not be able to disagree.
+	const profile = egressProfileLabel(egress)
+	const target = {
+		namespace: config.namespace,
+		name: egress.networkPolicyName ?? defaultEgressPolicyName(sandboxTemplateName, profile?.value),
 		sandboxTemplateName,
-	})
-	await verifyEgressPolicyApplied(client, translated, signal)
+		...(profile !== undefined ? { profile } : {}),
+	}
+	// Translated ONCE per boundary, not once per check: a `resolver` policy's
+	// `resolve()` is the host's own closure and may cost a network call, and
+	// running the two checks against two independently resolved allowlists
+	// would compare each against a different translation.
+	let translation: Promise<KubernetesTranslatedEgressPolicy> | undefined
+	const translate = (): Promise<KubernetesTranslatedEgressPolicy> => {
+		translation ??= translateEgressPolicy(
+			egress.policy,
+			engine,
+			target,
+			egress.ciliumNarrowing,
+		).catch((err: unknown) => {
+			translation = undefined
+			throw err
+		})
+		return translation
+	}
+	let namedObject: Promise<void> | undefined
+	const passes = new Map<string, CachedPass>()
+	// The per-sandbox selector label carries a once-ever value, so it is
+	// excluded from the memo key — see {@link policyCacheKey}. Resolved once
+	// here rather than per check, because the resolver validates as it
+	// resolves and a per-check throw would surface from a cache lookup.
+	const perSandboxKey = perSandboxEgressLabelKey(egress)
+	return {
+		async verifyNamedObject(signal) {
+			namedObject ??= (async () => {
+				await verifyEgressPolicyApplied(client, await translate(), signal)
+			})().catch((err: unknown) => {
+				namedObject = undefined
+				throw err
+			})
+			await namedObject
+		},
+		async verifyUnion(podLabels, subject, signal) {
+			if (!egressUnionVerificationEnabled(egress)) return
+			const translated = await translate()
+			const key = policyCacheKey(podLabels, perSandboxKey)
+			const cached = passes.get(key)
+			if (cached !== undefined && now() - cached.at < EGRESS_UNION_CACHE_TTL_MS) {
+				await cached.pending
+				return
+			}
+			const pending = verifyEgressPolicyUnion(
+				client,
+				translated,
+				{ namespace: config.namespace, podLabels, engine, subject },
+				signal,
+			).catch((err: unknown) => {
+				// A failed attempt is never cached — same rule the named-object
+				// memo has always had.
+				passes.delete(key)
+				throw err
+			})
+			passes.set(key, { at: now(), pending })
+			await pending
+		},
+	}
+}
+
+/**
+ * What a create path calls to prove the agent port is closed before it hands
+ * a sandbox back. `undefined` when `config.ingress` is `'unverified'`, so the
+ * opt-out is one absent function rather than a flag every call site re-reads.
+ */
+export type IngressVerifier = (
+	podLabels: Readonly<Record<string, string>>,
+	subject: string,
+	signal?: AbortSignal,
+) => Promise<void>
+
+/**
+ * Canonical key for one label set — order-independent, so two spellings of
+ * the same pod share a memo.
+ *
+ * `excludeKey` drops the PER-SANDBOX egress label, whose value is unique per
+ * acquire. Both memos exist to amortise a namespace-wide policy enumeration
+ * across every sandbox a backend produces, and a key that carried a
+ * once-ever value would give every acquire a miss and leave an entry behind
+ * that nothing ever looks up again — a full enumeration per sandbox, and a
+ * Map that grows for the host's whole life.
+ *
+ * Dropping it is sound at the moment these checks run: the only policy that
+ * could select a pod BY that key and value is that sandbox's own, whose name
+ * is generated in the same call and which does not exist yet. Every other
+ * policy selecting the pod — the operator's baseline, a per-profile one,
+ * anything hand-written — selects on the labels that remain, so two pods
+ * differing only in this label are the same question. The label itself is
+ * still PRESENT in the label set each check is run against; only the memo's
+ * key ignores it.
+ */
+function policyCacheKey(podLabels: Readonly<Record<string, string>>, excludeKey?: string): string {
+	return JSON.stringify(
+		Object.entries(podLabels)
+			.filter(([k]) => k !== excludeKey)
+			.sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0)),
+	)
+}
+
+/**
+ * Build the ingress check this config asks for, or nothing at all.
+ *
+ * `cache` is the provider path's per-label-set memo; a workspace passes none,
+ * and re-checks on every call for the same reason its egress check does —
+ * creating a workspace is a rare, explicit act with nothing to amortise, and
+ * a policy deleted since the last call must be noticed.
+ */
+export function buildIngressVerifier(
+	client: KubernetesClient,
+	config: KubernetesBackendInternalConfig,
+	cache?: Map<string, Promise<void>>,
+): IngressVerifier | undefined {
+	if (!ingressVerificationEnabled(config.ingress)) return undefined
+	const engine = resolveIngressEngine(config.ingress, config.egress?.engine)
+	const agentPort = config.agentPort ?? DEFAULT_AGENT_PORT
+	// Same exclusion, same reason, as the egress union memo above: a
+	// once-ever label value in the key would make this memo a per-acquire
+	// miss and an unbounded Map. See {@link policyCacheKey}.
+	const perSandboxKey = perSandboxEgressLabelKey(config.egress)
+	return async (podLabels, subject, signal) => {
+		const target = { namespace: config.namespace, podLabels, agentPort, engine, subject }
+		if (cache === undefined) {
+			await verifyIngressPolicyApplied(client, target, signal)
+			return
+		}
+		const key = policyCacheKey(podLabels, perSandboxKey)
+		let pending = cache.get(key)
+		if (pending === undefined) {
+			pending = verifyIngressPolicyApplied(client, target, signal).catch((err: unknown) => {
+				cache.delete(key)
+				throw err
+			})
+			cache.set(key, pending)
+		}
+		await pending
+	}
 }
 
 /** Config → the client's own access shape. Shared with `workspace.ts`. */
@@ -455,6 +901,450 @@ export function clientAccess(config: KubernetesBackendInternalConfig): Kubernete
 	}
 }
 
+/**
+ * Why an acquire was refused, in the terms an operator acts on rather than
+ * the terms the failure happened to arrive in.
+ *
+ *  - `'api-unreachable'` — the API server could not be reached, or kept
+ *    answering with a status that means "not now": a connect failure, a 429,
+ *    a 5xx. Retried inside the readiness budget before it ever reaches a
+ *    caller, so seeing it means the whole budget was spent failing.
+ *  - `'api-timeout'` — requests were accepted and never answered, until
+ *    `apiRequestTimeoutMs` gave up on them. Also retried first.
+ *  - `'forbidden'` — 401 or 403. The host's ServiceAccount cannot do this;
+ *    no amount of waiting changes that. See the RBAC section of
+ *    `docs/sdk/kubernetes-sandbox.md`.
+ *  - `'claim-rejected'` — the controller REFUSED the claim, and said why.
+ *    {@link KubernetesAcquireError.controllerReason} carries its own word for
+ *    it. This is the one that used to burn the entire readiness budget before
+ *    failing.
+ *  - `'capacity'` — the pod exists and cannot be placed: the scheduler
+ *    reports `PodScheduled=False` with reason `Unschedulable`. The cluster is
+ *    full, or nothing matches the template's placement rules.
+ *  - `'image-pull'` — the pod was placed and its container cannot start
+ *    because the image will not pull. Permanent until an operator fixes the
+ *    reference or the pull credential.
+ *  - `'not-ready'` — none of the above: the readiness budget expired with the
+ *    cluster reporting nothing wrong. A slow cold start, a webhook, an
+ *    admission controller, a CNI that never attached the pod.
+ */
+export type KubernetesAcquireFailureReason =
+	| 'api-unreachable'
+	| 'api-timeout'
+	| 'forbidden'
+	| 'claim-rejected'
+	| 'capacity'
+	| 'image-pull'
+	| 'not-ready'
+
+/**
+ * An acquire that was refused, carrying WHY in a field rather than in prose.
+ *
+ * Before this class a burst past node capacity and an API outage were the
+ * same plain `Error`, and a host could only tell them apart by matching
+ * message text that any release is free to reword. `reason` is the diagnosis,
+ * `retryable` is the advice that follows from it, and `cause` is the original
+ * failure — unmodified, so a host that already catches
+ * {@link ReadinessPollTimeout}, {@link KubernetesApiTimeoutError} or
+ * `KubernetesCredentialError` finds it there.
+ *
+ * `retryable` is about THIS acquire being worth attempting again, not about
+ * anything having been retried. Transient API failures are already retried
+ * inside the readiness budget, so a `retryable: true` that reaches a caller
+ * means the whole budget was spent on them.
+ *
+ * Not every acquire failure becomes one of these, and that is deliberate: a
+ * refusal this class cannot honestly diagnose — a malformed template, a 400
+ * from an admission webhook, a controller that reported Ready and named no
+ * sandbox — travels out as itself rather than being filed under whichever of
+ * the seven reasons is least wrong. `KubernetesApiError` carries the status
+ * for those.
+ */
+export class KubernetesAcquireError extends Error {
+	override readonly name = 'KubernetesAcquireError'
+	readonly reason: KubernetesAcquireFailureReason
+	/** Whether attempting the same acquire again could plausibly succeed. */
+	readonly retryable: boolean
+	/**
+	 * `status.conditions[Ready].reason`, verbatim, when the controller
+	 * refused the claim — `WarmPoolNotFound`, `TemplateNotFound`,
+	 * `InvalidMetadata`, `EnvVarsInjectionRejected`. Present only for
+	 * `'claim-rejected'`.
+	 */
+	readonly controllerReason?: string
+	/** The controller's own message for the same condition. */
+	readonly controllerMessage?: string
+
+	constructor(details: {
+		readonly reason: KubernetesAcquireFailureReason
+		readonly retryable: boolean
+		readonly message: string
+		readonly controllerReason?: string
+		readonly controllerMessage?: string
+		readonly cause?: unknown
+	}) {
+		super(details.message, details.cause !== undefined ? { cause: details.cause } : undefined)
+		this.reason = details.reason
+		this.retryable = details.retryable
+		if (details.controllerReason !== undefined) this.controllerReason = details.controllerReason
+		if (details.controllerMessage !== undefined) this.controllerMessage = details.controllerMessage
+	}
+}
+
+/**
+ * The `status.conditions[Ready].reason` values that mean the controller has
+ * DECIDED, so waiting is pointless.
+ *
+ * ## Where these strings came from
+ *
+ * Not from the issue that asked for this, and not from upstream source: this
+ * repo vendors none of agent-sandbox's Go, so a literal copied out of a
+ * changelog is a literal nobody here can check. Each of the four was produced
+ * against the deployed controller (kind v1.37.0, agent-sandbox v1.0.2,
+ * 2026-09-17) by making the claim it describes and reading the condition
+ * back:
+ *
+ * | reason | how it was produced | the controller's message |
+ * |---|---|---|
+ * | `WarmPoolNotFound` | claim at a pool that does not exist | `SandboxWarmPool "…" not found` |
+ * | `TemplateNotFound` | claim at a pool whose template does not exist | `SandboxTemplate "…" not found` |
+ * | `InvalidMetadata` | claim with an `additionalPodMetadata` label outside the allowed domains | `invalid additionalPodMetadata: …` |
+ * | `EnvVarsInjectionRejected` | claim with `spec.env` against a template that forbids injection | `environment variable injection rejected: …` |
+ *
+ * The transient reasons seen on the SAME cluster, which must NOT be in this
+ * set, were `DependenciesNotReady` (pod exists, still Pending) and
+ * `DependenciesReady` (the Ready=True reason).
+ *
+ * ## Why an unknown reason is not terminal
+ *
+ * A wrong literal here fails in one of two ways, and only one of them is
+ * recoverable. Too few entries: a rejected claim waits out the readiness
+ * budget, which is exactly the behaviour every release before this one had.
+ * Too many: an acquire that would have succeeded is refused on a guess. So
+ * the set is a closed list of measured strings and everything else falls
+ * through to the deadline.
+ */
+// Frozen because it is exported from the package root: an array handed to
+// every consumer is one a cast can push onto, and an entry added there would
+// change fail-fast for the whole process. The `readonly string[]` annotation
+// is deliberate rather than `as const` — `includes` on a literal tuple only
+// accepts the literals, and the whole point is to ask it about a reason no
+// one here has seen.
+export const TERMINAL_CLAIM_REASONS: readonly string[] = Object.freeze([
+	'WarmPoolNotFound',
+	'TemplateNotFound',
+	'InvalidMetadata',
+	'EnvVarsInjectionRejected',
+])
+
+/**
+ * The `Ready` condition a claim reports, whatever its status — the one
+ * {@link isConditionTrue} deliberately cannot return, because it answers a
+ * boolean question and this one needs the reason.
+ */
+function readyCondition(
+	conditions: readonly KubernetesCondition[] | undefined,
+): KubernetesCondition | undefined {
+	return conditions?.find((c) => c.type === READY_CONDITION)
+}
+
+/**
+ * What this backend asked the controller to put on the claim's pod, carried
+ * into the rejection so an `InvalidMetadata` refusal can say what was sent
+ * and which knob changes it.
+ *
+ * Context, never a second decision: whether a claim is refused at all is
+ * {@link TERMINAL_CLAIM_REASONS}' answer and nobody else's, so an empty map
+ * changes nothing about the class thrown, the reason on it, or when it is
+ * raised.
+ */
+export interface ClaimPodMetadataContext {
+	readonly namespace: string
+	/** `spec.additionalPodMetadata.labels`, exactly as sent. */
+	readonly requestedPodLabels: Readonly<Record<string, string>>
+	/** The egress profile among those labels, when one is configured. */
+	readonly profile?: EgressProfileLabel
+}
+
+/**
+ * A claim the controller has refused, or `undefined` for one it is still
+ * working on.
+ *
+ * `status: 'False'` alone is not a refusal — it is also what a claim looks
+ * like for the whole of a cold start — so the REASON decides, against
+ * {@link TERMINAL_CLAIM_REASONS}.
+ *
+ * ONE class comes out of here whatever the reason, and that is deliberate:
+ * `InvalidMetadata` is how the controller refuses a pod label whose domain is
+ * not on its allowlist — the profile's, today — and a host's `catch` must not
+ * have to be written differently depending on whether a profile happens to be
+ * configured. `metadata` only decides what rides along as the `cause`: a
+ * {@link KubernetesPodLabelsRejectedError} naming the map that was sent
+ * and the `config.egress.profileLabelKey` that moves it, which the
+ * controller's own message cannot know about.
+ */
+export function classifyClaimRejection(
+	claim: SandboxClaimResource | undefined,
+	claimName: string,
+	metadata?: ClaimPodMetadataContext,
+): KubernetesAcquireError | undefined {
+	const condition = readyCondition(claim?.status?.conditions)
+	if (condition === undefined || condition.status !== 'False') return undefined
+	const reason = condition.reason
+	if (reason === undefined || !TERMINAL_CLAIM_REASONS.includes(reason)) return undefined
+	// Only for the reason the pod metadata can actually cause, and only when
+	// this backend sent any: a `WarmPoolNotFound` carrying a labels-and-
+	// allowlist explanation would send an operator after the wrong thing.
+	const cause =
+		reason === 'InvalidMetadata' &&
+		metadata !== undefined &&
+		Object.keys(metadata.requestedPodLabels).length > 0
+			? new KubernetesPodLabelsRejectedError(
+					metadata.requestedPodLabels,
+					claimName,
+					metadata.namespace,
+					reason,
+					condition.message ?? '(the controller reported no message)',
+					metadata.profile,
+				)
+			: undefined
+	return new KubernetesAcquireError({
+		reason: 'claim-rejected',
+		retryable: false,
+		controllerReason: reason,
+		...(condition.message !== undefined ? { controllerMessage: condition.message } : {}),
+		...(cause !== undefined ? { cause } : {}),
+		message: `kubernetes: the agent-sandbox controller refused SandboxClaim ${claimName} with reason ${reason}${
+			condition.message !== undefined ? `: ${condition.message}` : ''
+		}. That is a decision, not a delay, so the readiness budget was not waited out.${
+			cause !== undefined ? ` ${cause.message}` : ''
+		}`,
+	})
+}
+
+/**
+ * How long to wait before repeating a failed readiness read, or `undefined`
+ * when the failure is not worth repeating.
+ *
+ * Retryable: a connect failure (the socket, not the answer), a request the
+ * `apiRequestTimeoutMs` bound gave up on, a 429 (the API server's own
+ * priority-and-fairness queue shedding load) and any 5xx. Not retryable: 401
+ * and 403, which are a decision; 404 and 410, which are an answer; 409, which
+ * a caller resolves by re-reading; and everything this backend threw itself.
+ *
+ * The wait is the poll's own cadence unless the server named one — then its
+ * `Retry-After`, because the server knows when its queue drains and this code
+ * does not. Nothing here consults a clock: the caller's deadline owns the
+ * sleep, so a long `Retry-After` spends the readiness budget rather than
+ * extending it.
+ */
+/**
+ * The largest delay a timer can hold — `2^31 - 1` ms, Node's own ceiling.
+ * Above it `setTimeout` warns and fires immediately, which is the opposite of
+ * what a long `Retry-After` asked for.
+ */
+const MAX_RETRY_DELAY_MS = 2_147_483_647
+
+export function retryDelayForApiFailure(err: unknown, pollIntervalMs: number): number | undefined {
+	if (err instanceof KubernetesApiTimeoutError) return pollIntervalMs
+	if (!(err instanceof KubernetesApiError)) return undefined
+	if (err.transport === 'connect') return pollIntervalMs
+	const status = err.status
+	if (status === undefined) return undefined
+	// `Retry-After` may only ever SLOW the poll down, and only within what a
+	// timer can express. A header of `0`, or one naming a moment already past,
+	// would otherwise turn the retry into a hot loop against a server that is
+	// already shedding load — the caller's own cadence is the rate this loop
+	// runs at when nothing is wrong. And a header naming a moment years away
+	// overflows `setTimeout`, which then fires at once rather than never,
+	// producing the same hot loop from the opposite direction. The readiness
+	// deadline ends the wait either way; the clamp only stops the wait from
+	// silently becoming no wait at all.
+	if (status === 429 || status >= 500) {
+		const asked = err.retryAfterMs ?? pollIntervalMs
+		return Math.min(Math.max(asked, pollIntervalMs), MAX_RETRY_DELAY_MS)
+	}
+	return undefined
+}
+
+/**
+ * How long the one diagnostic pod read after a failed acquire may take.
+ *
+ * Same shape and the same argument as `runFailureCleanup`'s grace: the
+ * readiness clock has already expired, so this cannot share it, and a
+ * diagnosis that could hang would keep `create()` pending past the budget the
+ * caller chose — for a nicer error message. One second, and a diagnosis that
+ * does not arrive is simply not made.
+ */
+const ACQUIRE_DIAGNOSIS_GRACE_MS = 1_000
+
+/**
+ * The image-pull `status.containerStatuses[].state.waiting.reason` values the
+ * kubelet reports. Measured on kind v1.37.0 (2026-09-17): a container whose
+ * image does not exist waits as `ErrImagePull` for the first attempts and
+ * settles into `ImagePullBackOff`. The other two are the kubelet's names for
+ * a pull that resolved and then failed, and for a registry that cannot be
+ * reached at all.
+ */
+const IMAGE_PULL_WAITING_REASONS: readonly string[] = [
+	'ErrImagePull',
+	'ImagePullBackOff',
+	'ImageInspectError',
+	'RegistryUnavailable',
+]
+
+/**
+ * Ask the pod why it is not ready, once, after the budget has already gone.
+ *
+ * Nothing on the healthy path calls this and nothing waits on it: it runs
+ * exactly when an acquire has already failed, and its whole output is a
+ * better {@link KubernetesAcquireFailureReason} than `'not-ready'`. A read
+ * that fails, a pod that is not there and a pod with nothing to say all
+ * produce `undefined`, which leaves the reason where it was.
+ *
+ * It must run BEFORE the cleanup DELETE, because the pod goes away with the
+ * object it belongs to.
+ *
+ * It takes the CALLER's signal where `runFailureCleanup` deliberately does
+ * not: cleanup must finish or the cluster keeps the object, while a diagnosis
+ * is only a better sentence for an error a caller who aborted will never
+ * read.
+ */
+async function diagnoseUnreadyPod(
+	client: KubernetesClient,
+	namespace: string,
+	podName: string,
+	callerSignal: AbortSignal | undefined,
+): Promise<'capacity' | 'image-pull' | undefined> {
+	let pod: PodResource | undefined
+	try {
+		const deadline = new OperationDeadline(
+			ACQUIRE_DIAGNOSIS_GRACE_MS,
+			'kubernetes acquire diagnosis',
+			callerSignal,
+		)
+		pod = await deadline.run((signal) =>
+			client.request<PodResource>('GET', podPath(namespace, podName), undefined, signal),
+		)
+	} catch {
+		// The acquire failure is the primary one and keeps its reason. A
+		// diagnosis that cannot be made is not a second failure to report.
+		return undefined
+	}
+	const scheduled = pod?.status?.conditions?.find((c) => c.type === 'PodScheduled')
+	if (scheduled?.status === 'False' && scheduled.reason === 'Unschedulable') return 'capacity'
+	for (const container of pod?.status?.containerStatuses ?? []) {
+		const reason = container.state?.waiting?.reason
+		if (reason !== undefined && IMAGE_PULL_WAITING_REASONS.includes(reason)) return 'image-pull'
+	}
+	return undefined
+}
+
+/**
+ * The refusal a caller sees, given the failure that actually happened and
+ * whatever the pod had to say about it.
+ *
+ * Returns `undefined` for a failure none of the seven reasons describes —
+ * see {@link KubernetesAcquireError} for why that is a deliberate hole rather
+ * than a missing case. A {@link KubernetesAcquireError} that arrived from
+ * deeper in (the claim rejection) is returned unchanged: it is already the
+ * diagnosis.
+ */
+export function classifyAcquireFailure(
+	err: unknown,
+	podDiagnosis: 'capacity' | 'image-pull' | undefined,
+): KubernetesAcquireError | undefined {
+	if (err instanceof KubernetesAcquireError) return err
+	if (err instanceof KubernetesApiTimeoutError) {
+		return new KubernetesAcquireError({
+			reason: 'api-timeout',
+			retryable: true,
+			message: `kubernetes: the acquire was refused because the API server did not answer in time — ${err.message}`,
+			cause: err,
+		})
+	}
+	if (err instanceof KubernetesCredentialError) {
+		return new KubernetesAcquireError({
+			reason: 'forbidden',
+			retryable: false,
+			message: `kubernetes: the acquire was refused because the API server rejected this host's credential — ${err.message}. Check the host ServiceAccount's Role against the RBAC section of the Kubernetes sandbox documentation.`,
+			cause: err,
+		})
+	}
+	if (err instanceof KubernetesApiError) {
+		// The same predicate the poll retries on, so "worth trying again" has
+		// one definition and a caller cannot be told a failure is retryable
+		// that the poll would have declined to retry. The interval is
+		// irrelevant here — only whether an answer comes back at all.
+		if (retryDelayForApiFailure(err, 1) === undefined) return undefined
+		return new KubernetesAcquireError({
+			reason: 'api-unreachable',
+			retryable: true,
+			message: `kubernetes: the acquire was refused because the API server could not serve it — ${err.message}`,
+			cause: err,
+		})
+	}
+	if (err instanceof ReadinessPollTimeout) {
+		// ORDER MATTERS, and this is the order: what the cluster SAID beats
+		// what the failures suggest. A pod diagnosis is a condition the API
+		// server published about this pod, read after the budget had already
+		// gone; `err.cause` is at best the failure the poll was still meeting
+		// at that moment. On a saturated cluster both are present at once — a
+		// pod nothing can schedule AND an API server shedding load — and
+		// reporting `api-unreachable` there would hide the very reason this
+		// function exists to produce, and would turn `image-pull`'s
+		// `retryable: false` into a `true` that has a host retrying forever
+		// against an image reference that will never resolve. A diagnosis also
+		// cannot be stale in the way a cause can: it only exists because the
+		// API server answered one more read, moments ago.
+		if (podDiagnosis === 'capacity') {
+			return new KubernetesAcquireError({
+				reason: 'capacity',
+				retryable: true,
+				message: `kubernetes: the acquire was refused because its pod could not be scheduled — the cluster reports PodScheduled=False/Unschedulable. ${err.message}`,
+				cause: err,
+			})
+		}
+		if (podDiagnosis === 'image-pull') {
+			return new KubernetesAcquireError({
+				reason: 'image-pull',
+				retryable: false,
+				message: `kubernetes: the acquire was refused because its pod's container image will not pull. ${err.message}`,
+				cause: err,
+			})
+		}
+		// Nothing measured, so the failures the poll kept meeting decide: a
+		// poll that spent its budget retrying API failures did not fail
+		// because the sandbox was slow; it failed because the control plane
+		// was. `pollForBinding` carries the failure it was STILL meeting onto
+		// the timeout, and the reason follows it rather than the timeout.
+		const underlying = classifyAcquireFailure(err.cause, undefined)
+		if (underlying !== undefined) {
+			return new KubernetesAcquireError({
+				reason: underlying.reason,
+				retryable: underlying.retryable,
+				message: underlying.message,
+				cause: err,
+			})
+		}
+		return new KubernetesAcquireError({
+			reason: 'not-ready',
+			retryable: true,
+			message: err.message,
+			cause: err,
+		})
+	}
+	if (err instanceof OperationDeadlineExpired) {
+		return new KubernetesAcquireError({
+			reason: 'not-ready',
+			retryable: true,
+			message: `kubernetes: the acquire ran out of readiness budget — ${err.message}`,
+			cause: err,
+		})
+	}
+	return undefined
+}
+
 /**
  * Claim or create, wait for Ready, read the bound identity back, resolve the
  * address and learn the pod's uid — or leave nothing behind trying.
@@ -462,12 +1352,39 @@ export function clientAccess(config: KubernetesBackendInternalConfig): Kubernete
  * Exported because the sandbox surface is built on top of this record rather
  * than beside it: one acquire path, one cleanup path, whatever ends up
  * wrapping them.
+ *
+ * ## What it refuses with
+ *
+ * Every refusal this function can diagnose arrives as a
+ * {@link KubernetesAcquireError} naming one of seven reasons, with the
+ * original failure as its `cause`. Three things stay outside that:
+ * configuration refused before anything is created
+ * ({@link assertEnforceable}, {@link assertRuntimeClassIsApplicable}), a
+ * caller's own abort, and a failure none of the seven reasons honestly
+ * describes — see {@link KubernetesAcquireError} for why the last one is a
+ * hole on purpose.
+ *
+ * ## What it retries, and what it will not
+ *
+ * A readiness GET that fails transiently — a connect failure, a request the
+ * API bound gave up on, a 429, a 5xx — is repeated INSIDE the readiness
+ * deadline, honouring `Retry-After`. One clock, so a retry spends the budget
+ * rather than extending it, and a `create()` cannot outlive the timeout its
+ * caller chose. The create POST is never retried: it is not idempotent, and a
+ * POST whose answer never arrived may already have committed — which is why
+ * cleanup deletes the client-owned name whatever happened.
  */
 export async function acquireKubernetesSandbox(
 	client: KubernetesClient,
 	config: KubernetesBackendInternalConfig,
 	options: SandboxBackendOptions,
 	readiness: { readonly timeoutMs: number; readonly pollIntervalMs: number },
+	verifyIngress: IngressVerifier | undefined = buildIngressVerifier(client, config),
+	egressBoundary: KubernetesEgressBoundary | undefined = buildEgressBoundary(
+		client,
+		config,
+		config.sandboxTemplateName,
+	),
 ): Promise<KubernetesAcquisition> {
 	options.signal?.throwIfAborted()
 	assertEnforceable(options)
@@ -527,70 +1444,241 @@ export async function acquireKubernetesSandbox(
 		config.warmPoolName !== undefined
 			? claimCollectionPath(namespace)
 			: sandboxCollectionPath(namespace)
-	const createBody =
-		config.warmPoolName !== undefined
-			? buildClaimBody(namespace, objectName, config.warmPoolName, shutdownTime)
-			: buildSandboxBody({
-					namespace,
-					name: objectName,
-					template: await deadline.run((signal) =>
-						readSandboxTemplate(client, namespace, config.sandboxTemplateName, signal),
-					),
-					sandboxTemplateName: config.sandboxTemplateName,
-					shutdownTime,
-					...(config.runtimeClassName !== undefined
-						? { runtimeClassName: config.runtimeClassName }
-						: {}),
-				})
+	// Composed ONCE, here, and handed to whichever body builder runs below:
+	// the claim's `additionalPodMetadata.labels` and a direct Sandbox's pod
+	// template metadata are the same map, and the translated policy's selector
+	// is built from the same resolution. See `composeAdditionalPodLabels`.
+	// Empty — the only case before a profile is configured — means every body
+	// below is byte for byte what it was.
+	// The per-sandbox selector label rides in the SAME map, through the same
+	// composer, as a second key rather than a second construction — the whole
+	// reason `composeAdditionalPodLabels` takes an `extra`. Its value is the
+	// name of the object this acquire is about to create, which is unique per
+	// acquire and known BEFORE the POST: that is what lets the label travel as
+	// claim-time pod metadata (warm-safe, no cold start) instead of as a patch
+	// to a running pod this backend has no verb for.
+	const perSandboxLabelKey = perSandboxEgressLabelKey(config.egress)
+	const podLabels = composeAdditionalPodLabels(
+		config.egress,
+		perSandboxLabelKey !== undefined ? { [perSandboxLabelKey]: objectName } : undefined,
+	)
+	const profile = egressProfileLabel(config.egress)
+	// Read off the create reply, and off the readiness polls if that reply
+	// carried no object — it is the uid a per-sandbox policy's
+	// `ownerReferences` names and the suffix of its name, so the cluster can
+	// garbage-collect the policy with the object this backend owns.
+	let ownerUid: string | undefined
+	const buildCreateBody = async (): Promise<Record<string, unknown>> => {
+		if (config.warmPoolName !== undefined) {
+			return buildClaimBody({
+				namespace,
+				name: objectName,
+				warmPoolName: config.warmPoolName,
+				shutdownTime,
+				...(config.claimLabels !== undefined ? { labels: config.claimLabels } : {}),
+				podLabels,
+			})
+		}
+		const template = await deadline.run((signal) =>
+			readSandboxTemplate(client, namespace, config.sandboxTemplateName, signal),
+		)
+		// A direct Sandbox's pod labels are decided HERE, by the body below,
+		// so both network boundaries are checked against the real labels before
+		// anything is created — a refusal leaves no Sandbox and no PVC behind
+		// rather than one of each to clean up. This is EARLIER than the plan
+		// for the egress union check asked for (it said "after binding", for
+		// the bound pod's labels); a pool-less Sandbox's labels are knowable
+		// before the POST, and refusing with nothing created is strictly
+		// better than refusing with an object to clean up.
+		const directPodLabels = sandboxPodLabels(template, config.sandboxTemplateName, podLabels)
+		const directSubject = `to create Sandbox ${objectName} in namespace ${namespace}`
+		if (verifyIngress !== undefined) {
+			await deadline.run((signal) => verifyIngress(directPodLabels, directSubject, signal))
+		}
+		if (egressBoundary !== undefined) {
+			await deadline.run((signal) =>
+				egressBoundary.verifyUnion(directPodLabels, directSubject, signal),
+			)
+		}
+		return buildSandboxBody({
+			namespace,
+			name: objectName,
+			template,
+			sandboxTemplateName: config.sandboxTemplateName,
+			shutdownTime,
+			podLabels,
+			...(config.runtimeClassName !== undefined
+				? { runtimeClassName: config.runtimeClassName }
+				: {}),
+		})
+	}
+
+	let createBody: Record<string, unknown>
+	try {
+		createBody = await buildCreateBody()
+	} catch (err) {
+		// Nothing exists yet, so there is nothing to clean up and no pod to
+		// ask — but a 403 on the template read is still a `forbidden` acquire,
+		// and a caller should not have to tell that apart by where it
+		// happened.
+		throw classifyAcquireFailure(err, undefined) ?? err
+	}
+
+	// The pod the diagnosis below asks, when there is one. A directly created
+	// Sandbox is backed by a pod of its own name; a CLAIM's pod is not knowable
+	// until the controller has named a sandbox in `status.sandbox`, which it
+	// does before Ready on a cold start and never on a rejected claim.
+	let diagnosablePodName: string | undefined =
+		config.warmPoolName !== undefined ? undefined : objectName
+	// One definition of "worth trying again", shared by the poll that retries
+	// and the classification that reports — see {@link retryDelayForApiFailure}.
+	const pollBehaviour: ReadinessPollBehaviour = {
+		retryDelayFor: (err) => retryDelayForApiFailure(err, readiness.pollIntervalMs),
+	}
 
 	try {
 		// Inside the cleanup block: a POST that fails client-side may still have
 		// committed, so the only safe assumption is that the object exists.
-		await deadline.run((signal) => client.request('POST', createPath, createBody, signal))
+		const created = await deadline.run((signal) =>
+			client.request<{ readonly metadata?: { readonly uid?: string } }>(
+				'POST',
+				createPath,
+				createBody,
+				signal,
+			),
+		)
+		ownerUid ??= created?.metadata?.uid
 		const binding =
 			config.warmPoolName !== undefined
 				? await pollForBinding(
-						async (signal) =>
-							bindingFromClaim(
-								await client.request<SandboxClaimResource>(
-									'GET',
-									claimPath(namespace, objectName),
-									undefined,
-									signal,
-								),
-								objectName,
-							),
+						async (signal) => {
+							const claim = await client.request<SandboxClaimResource>(
+								'GET',
+								claimPath(namespace, objectName),
+								undefined,
+								signal,
+							)
+							diagnosablePodName = claim?.status?.sandbox?.name ?? diagnosablePodName
+							ownerUid ??= claim?.metadata?.uid
+							// The pod labels this backend asked the controller for
+							// travel into the read, not because the fail-fast needs
+							// them — `InvalidMetadata` is already one of
+							// `TERMINAL_CLAIM_REASONS`, and that is the ONE
+							// taxonomy a refused claim is reported under — but so
+							// the refusal can carry what was actually sent and how
+							// to change it. See {@link ClaimPodMetadataContext}.
+							return bindingFromClaim(claim, objectName, {
+								namespace,
+								requestedPodLabels: podLabels,
+								...(profile !== undefined ? { profile } : {}),
+							})
+						},
 						deadline,
 						readiness,
 						`claim ${objectName}`,
+						pollBehaviour,
 					)
 				: await pollForBinding(
-						async (signal) =>
-							bindingFromSandbox(
-								await client.request<SandboxResource>(
-									'GET',
-									sandboxPath(namespace, objectName),
-									undefined,
-									signal,
-								),
-							),
+						async (signal) => {
+							const sandbox = await client.request<SandboxResource>(
+								'GET',
+								sandboxPath(namespace, objectName),
+								undefined,
+								signal,
+							)
+							ownerUid ??= sandbox?.metadata?.uid
+							return bindingFromSandbox(sandbox)
+						},
 						deadline,
 						readiness,
 						`sandbox ${objectName}`,
+						pollBehaviour,
 					)
 
-		const token = await deadline.run((signal) =>
-			readPodBindToken(client, namespace, binding, signal),
+		const agentPort = config.agentPort ?? DEFAULT_AGENT_PORT
+		const mode = config.agentAddress ?? 'service'
+		// One read, two facts: the bind token and — under `'pod-ip'` — the
+		// address, off the same pod. See {@link readAddressedPod}, which under
+		// a configured profile also waits for the label the controller
+		// patches onto the bound pod before anything is admitted.
+		const pod = await readAddressedPod(
+			client,
+			namespace,
+			binding,
+			deadline,
+			readiness,
+			mode,
+			podLabels,
 		)
+		// The one refusal this capability must not skip: an unlabelled pod
+		// handed back runs under whatever policy DOES select it while the host
+		// believes it is on a narrower profile. This throws INSIDE the try
+		// block, so the cleanup below releases the claim (or deletes the
+		// Sandbox) exactly as a failed privilege probe does — and the probe
+		// itself, which runs in `admitProbedSandbox` after this function
+		// returns, is therefore never reached with the label unobserved.
+		assertRequestedPodLabelsObserved(
+			pod,
+			podLabels,
+			`Sandbox ${binding.name} in namespace ${namespace}`,
+		)
+		// A CLAIMED sandbox's pod was built from the pool's own template, so
+		// its labels are not knowable until the controller has bound one.
+		// Checking here rather than not at all is the trade: a refusal
+		// releases the claim through the cleanup below, which is the same
+		// path a failed privilege probe takes.
+		if (config.warmPoolName !== undefined) {
+			const boundSubject = `the pod bound to Sandbox ${binding.name} in namespace ${namespace}`
+			if (verifyIngress !== undefined) {
+				await deadline.run((signal) => verifyIngress(pod.labels ?? {}, boundSubject, signal))
+			}
+			// The egress union check asks the same question of the same labels
+			// — which policies select THIS pod — so it runs at the same two
+			// points, and a refusal here releases the claim through the cleanup
+			// below, exactly as a failed privilege probe does.
+			if (egressBoundary !== undefined) {
+				await deadline.run((signal) =>
+					egressBoundary.verifyUnion(pod.labels ?? {}, boundSubject, signal),
+				)
+			}
+		}
+		// Only when the capability is configured, and only after the pod has
+		// been confirmed to carry the selector label: a policy written for a
+		// pod that never got the label would select nothing while the caller
+		// was told its egress had been narrowed.
+		const owner =
+			perSandboxLabelKey === undefined
+				? undefined
+				: {
+						kind: (config.warmPoolName !== undefined ? 'SandboxClaim' : 'Sandbox') as
+							| 'SandboxClaim'
+							| 'Sandbox',
+						name: objectName,
+						uid: assertOwnerUid(ownerUid, objectName, namespace),
+					}
 		return {
 			binding,
-			agent: resolveAgentAddress(binding, config.agentPort ?? DEFAULT_AGENT_PORT, token),
+			agent: resolveAgentAddress(binding, agentPort, pod.uid, {
+				mode,
+				...(pod.podIP !== undefined ? { podIP: pod.podIP } : {}),
+			}),
+			// Only the literal-address mode gets one — see the field.
+			...(mode === 'pod-ip'
+				? { refreshAgent: buildAgentAddressRefresh(client, namespace, binding, agentPort, mode) }
+				: {}),
 			ownedPath,
+			...(owner !== undefined ? { owner, perSandboxLabelValue: objectName } : {}),
 			ttlSeconds,
 			release,
 			renew,
 		}
 	} catch (err) {
+		// Asked BEFORE cleanup, because the pod goes away with the object, and
+		// only for a timeout — every other failure already knows what it was.
+		const podDiagnosis =
+			err instanceof ReadinessPollTimeout && diagnosablePodName !== undefined
+				? await diagnoseUnreadyPod(client, namespace, diagnosablePodName, options.signal)
+				: undefined
 		// One cleanup for every way out of the block above, on its own short
 		// budget: the readiness clock has already expired in the common case,
 		// so spending it again would either skip cleanup or leave `create()`
@@ -598,35 +1686,258 @@ export async function acquireKubernetesSandbox(
 		await runFailureCleanup(async (signal) => {
 			await release(signal)
 		})
-		throw err
+		throw classifyAcquireFailure(err, podDiagnosis) ?? err
 	}
 }
 
 /**
- * The claim body, in full. Everything absent from it is absent on purpose:
- * no `env`, no `volumeClaimTemplates`, no `additionalPodMetadata`. Either of
- * the first two forces a cold start upstream and takes the warm pool away.
+ * The claim body, in full. Everything absent from it is absent on purpose: no
+ * `env` and no `volumeClaimTemplates`, because either forces a cold start
+ * upstream and takes the warm pool away. `additionalPodMetadata` is the one
+ * piece of claim-time metadata that does NOT — see `podLabels` below.
  *
  * `shutdownTime` + `shutdownPolicy: 'Delete'` is the leak guard: it bounds the
  * object by the wall clock whatever the host does, so a host that dies
  * mid-acquire costs the cluster one TTL rather than one leaked sandbox
  * forever. `ttlSecondsAfterFinished` deliberately does NOT appear — its timer
  * starts from the Finished condition, which a crashed host never reaches.
+ *
+ * `labels` (from {@link KubernetesBackendInternalConfig.claimLabels}) is the
+ * host's own bookkeeping and goes ONLY onto `metadata.labels` — never into
+ * `additionalPodMetadata`, because those are POD labels that change what
+ * selectors match a running sandbox, and a host's crash-recovery identity has
+ * no business doing that. Absent or empty, the body is exactly what it was
+ * before `claimLabels` existed.
+ *
+ * `podLabels` is the other map, on the other object: whatever
+ * `composeAdditionalPodLabels` produced — the egress profile today — which
+ * the controller merges onto the pod it binds. It is the one piece of
+ * claim-time metadata that does NOT cost a cold start, which is why `env` and
+ * `volumeClaimTemplates` are still absent from this body and this is not.
+ * Empty, `additionalPodMetadata` does not appear at all and the body is byte
+ * for byte what it always was.
  */
-function buildClaimBody(
-	namespace: string,
-	name: string,
-	warmPoolName: string,
-	shutdownTime: string,
-): Record<string, unknown> {
+interface ClaimBodyOptions {
+	readonly namespace: string
+	readonly name: string
+	readonly warmPoolName: string
+	readonly shutdownTime: string
+	/** `metadata.labels` on the CLAIM. See above. */
+	readonly labels?: Record<string, string>
+	/** `spec.additionalPodMetadata.labels` — the POD's. See above. */
+	readonly podLabels?: Readonly<Record<string, string>>
+}
+
+function buildClaimBody(options: ClaimBodyOptions): Record<string, unknown> {
+	const { labels, podLabels } = options
 	return {
 		apiVersion: `${SANDBOX_EXTENSIONS_API_GROUP}/${SANDBOX_API_VERSION}`,
 		kind: 'SandboxClaim',
-		metadata: { name, namespace },
+		metadata: {
+			name: options.name,
+			namespace: options.namespace,
+			...(labels !== undefined && Object.keys(labels).length > 0 ? { labels } : {}),
+		},
 		spec: {
-			warmPoolRef: { name: warmPoolName },
-			lifecycle: { shutdownTime, shutdownPolicy: 'Delete' },
+			warmPoolRef: { name: options.warmPoolName },
+			...(podLabels !== undefined && Object.keys(podLabels).length > 0
+				? { additionalPodMetadata: { labels: podLabels } }
+				: {}),
+			lifecycle: { shutdownTime: options.shutdownTime, shutdownPolicy: 'Delete' },
+		},
+	}
+}
+
+/**
+ * Refuse a bound pod that does not carry EVERY label this backend asked the
+ * controller to put on it — the egress profile today, and whatever else
+ * `composeAdditionalPodLabels` later contributes to the same map.
+ *
+ * The same set {@link readAddressedPod} waits for, deliberately: a wait that
+ * covered more than the refusal would burn the readiness budget on a label
+ * nothing then checked, and a refusal that covered more than the wait would
+ * refuse a pod that had simply not been patched yet. An empty map (the
+ * unprofiled path) checks nothing and returns.
+ *
+ * Called after {@link readAddressedPod} has already waited on the readiness
+ * deadline, so reaching a missing label here means it never arrived, not that
+ * it had not arrived yet.
+ */
+function assertRequestedPodLabelsObserved(
+	pod: KubernetesBoundPod,
+	requested: Readonly<Record<string, string>>,
+	subject: string,
+): void {
+	for (const [key, value] of Object.entries(requested)) {
+		if (pod.labels?.[key] === value) continue
+		throw new KubernetesPodLabelNotObservedError({ key, value }, subject, pod.labels ?? {})
+	}
+}
+
+/**
+ * The uid of the object this acquire created, or a refusal naming what was
+ * missing.
+ *
+ * Only reached when `config.egress.perSandbox` is configured, and only after
+ * a create reply and every readiness poll have been read without one — the
+ * API server assigns `metadata.uid` on admission and returns the object it
+ * created, so an object with no uid is an API server that answered something
+ * other than what it was asked for. Refusing is right: without the uid there
+ * is no owner reference, and a per-sandbox policy with no owner is one the
+ * cluster never collects.
+ */
+function assertOwnerUid(uid: string | undefined, objectName: string, namespace: string): string {
+	if (uid !== undefined && uid !== '') return uid
+	throw new KubernetesOwnerUidMissingError(objectName, namespace)
+}
+
+/**
+ * Options for {@link releaseKubernetesTaskSandboxes}.
+ */
+export interface KubernetesReleaseTaskSandboxesOptions {
+	/**
+	 * Required, and refused if empty — see {@link releaseKubernetesTaskSandboxes}.
+	 * The same selector syntax a `kubectl get --selector` takes, e.g.
+	 * `sandbox.namzu.ai/host-instance=host-a`.
+	 */
+	readonly labelSelector: string
+	readonly signal?: AbortSignal
+}
+
+/**
+ * Recover a crashed host's claims: LIST every `SandboxClaim` carrying
+ * `labelSelector`, `DELETE` each, and report what was removed.
+ *
+ * This deletes CLAIMS only. The controller's own garbage collection —
+ * ownerReferences from claim to the Sandbox it bound, and from Sandbox to
+ * Pod and Service — takes the rest down behind it; nothing here reads or
+ * touches a Sandbox or a Pod directly. A claim already gone (raced by the
+ * controller's own TTL reaper, or a second release call) counts as removed
+ * rather than a failure, the same convention every other DELETE in this
+ * backend follows.
+ *
+ * `labelSelector` is REQUIRED and refused, synchronously, before a single
+ * request goes out, if it is absent or empty: a release that could fall back
+ * to matching every claim (or every claim of the template) would delete a
+ * live fleet's work the first time a caller passed one by mistake. There is
+ * no default selector for exactly this reason.
+ */
+export async function releaseKubernetesTaskSandboxes(
+	config: KubernetesBackendInternalConfig,
+	options: KubernetesReleaseTaskSandboxesOptions,
+): Promise<{ readonly deleted: number; readonly names: readonly string[] }> {
+	if (options.labelSelector === '') {
+		throw new Error(
+			'kubernetes: releaseKubernetesTaskSandboxes requires a non-empty labelSelector — a release with no selector would delete every SandboxClaim in the namespace, including ones a live host still owns. Pass the selector that names only the claims you mean to recover.',
+		)
+	}
+	options.signal?.throwIfAborted()
+	const namespace = config.namespace
+	const client = createKubernetesClient(clientAccess(config), clientOptions(config))
+	const list = await client.request<SandboxClaimListResource>(
+		'GET',
+		claimListPath(namespace, options.labelSelector),
+		undefined,
+		options.signal,
+	)
+	const names = (list?.items ?? [])
+		.map((claim) => claim.metadata?.name)
+		.filter((name): name is string => typeof name === 'string' && name !== '')
+	// Concurrent, not one at a time: a crash-recovery release can carry a
+	// whole host's worth of claims, and nothing here needs the ordering a
+	// sequential loop would impose — each DELETE is independent and
+	// idempotent (an already-gone claim is tolerated below). A non-tolerated
+	// failure still rejects the whole call, exactly as a sequential loop
+	// would have on its first such failure.
+	await Promise.all(
+		names.map(async (name) => {
+			try {
+				await client.request('DELETE', claimPath(namespace, name), undefined, options.signal)
+			} catch (err) {
+				if (!(err instanceof KubernetesAlreadyGoneError)) throw err
+			}
+		}),
+	)
+	return { deleted: names.length, names }
+}
+
+/** Result of {@link readKubernetesTaskCapacity}. */
+export interface KubernetesTaskCapacity {
+	/** `config.warmPoolName`'s own replica counts, straight off its `status`/`spec`. */
+	readonly warmPool: {
+		/** `status.readyReplicas`. `0` when the field is absent (a brand-new or empty pool). */
+		readonly ready: number
+		/** `spec.replicas`. `0` when the field is absent. */
+		readonly desired: number
+	}
+	/** Every `SandboxClaim` in the namespace bound to `config.warmPoolName`, whatever its state. */
+	readonly activeClaims: number
+	/** Every Pod in the namespace currently in phase `Pending`. */
+	readonly pendingPods: number
+}
+
+/** Options for {@link readKubernetesTaskCapacity}. */
+export interface KubernetesReadTaskCapacityOptions {
+	readonly signal?: AbortSignal
+}
+
+/**
+ * Read task-pool headroom before admitting more work: three GETs, no writes.
+ *
+ * `config.warmPoolName` is required — this reads the exact object a claim's
+ * `warmPoolRef` names, so a pool-less backend (every create is a direct
+ * Sandbox) has no pool to report on. `activeClaims` is every claim in the
+ * namespace whose `spec.warmPoolRef.name` matches this pool, counted rather
+ * than trusted from a label, because a claim's `warmPoolRef` is the one
+ * field the API itself guarantees. `pendingPods` is every Pod in the
+ * namespace still in phase `Pending` — a coarse but honest signal of
+ * in-flight scale-up the ready-replica count alone does not carry, on
+ * either the warm or the pool-less path.
+ */
+export async function readKubernetesTaskCapacity(
+	config: KubernetesBackendInternalConfig,
+	options?: KubernetesReadTaskCapacityOptions,
+): Promise<KubernetesTaskCapacity> {
+	options?.signal?.throwIfAborted()
+	if (config.warmPoolName === undefined) {
+		throw new Error(
+			'kubernetes: readKubernetesTaskCapacity requires config.warmPoolName — there is no SandboxWarmPool to report on for a backend that creates every sandbox directly.',
+		)
+	}
+	const namespace = config.namespace
+	const warmPoolName = config.warmPoolName
+	const client = createKubernetesClient(clientAccess(config), clientOptions(config))
+	const [pool, claims, pods] = await Promise.all([
+		client.request<SandboxWarmPoolResource>(
+			'GET',
+			warmPoolPath(namespace, warmPoolName),
+			undefined,
+			options?.signal,
+		),
+		client.request<SandboxClaimListResource>(
+			'GET',
+			claimCollectionPath(namespace),
+			undefined,
+			options?.signal,
+		),
+		client.request<PodListResource>(
+			'GET',
+			podCollectionPath(namespace),
+			undefined,
+			options?.signal,
+		),
+	])
+	const activeClaims = (claims?.items ?? []).filter(
+		(claim) => claim.spec?.warmPoolRef?.name === warmPoolName,
+	).length
+	const pendingPods = (pods?.items ?? []).filter((pod) => pod.status?.phase === 'Pending').length
+	return {
+		warmPool: {
+			ready: pool?.status?.readyReplicas ?? 0,
+			desired: pool?.spec?.replicas ?? 0,
 		},
+		activeClaims,
+		pendingPods,
 	}
 }
 
@@ -649,6 +1960,34 @@ export interface SandboxBodyOptions {
 	 * sets it, because an unbounded task sandbox is a leak.
 	 */
 	readonly shutdownTime?: string
+	/**
+	 * Annotations to stamp on the Sandbox's OWN metadata at creation.
+	 *
+	 * One caller, and everything it writes is a fact the object has to carry
+	 * from the moment it exists rather than from its first patch: the holder
+	 * epoch of a workspace created under one, so there is no window in which
+	 * it stands unfenced, and the revision of the pod template it was built
+	 * from, so there is no window in which it claims none. Both are
+	 * `workspace.ts`'s — see `HOLDER_EPOCH_ANNOTATION_KEY` and
+	 * `POD_TEMPLATE_HASH_ANNOTATION_KEY`.
+	 *
+	 * Absent, the body is byte for byte what it always was, which is what
+	 * keeps every task sandbox's create unchanged — a task sandbox is
+	 * ephemeral, so it has no revision to drift from and nothing to fence.
+	 */
+	readonly annotations?: Readonly<Record<string, string>>
+	/**
+	 * Extra labels stamped onto the POD template's metadata, beside the
+	 * template label this body always adds.
+	 *
+	 * The same map `buildClaimBody` puts on a claim's
+	 * `additionalPodMetadata.labels`, from the same
+	 * `composeAdditionalPodLabels` call — a direct Sandbox has no controller
+	 * to merge them for it, so the create body stamps them itself and the two
+	 * paths produce one set of pod labels. Absent or empty, the pod template
+	 * is byte for byte what it was.
+	 */
+	readonly podLabels?: Readonly<Record<string, string>>
 }
 
 /**
@@ -681,22 +2020,14 @@ export interface SandboxBodyOptions {
  * label the translated policy is built to match.
  */
 export function buildSandboxBody(options: SandboxBodyOptions): Record<string, unknown> {
-	const podTemplate = options.template.podTemplate
-	const spec =
-		options.runtimeClassName !== undefined
-			? { ...podTemplate.spec, runtimeClassName: options.runtimeClassName }
-			: { ...podTemplate.spec }
-	const metadata = {
-		...podTemplate.metadata,
-		labels: {
-			...podTemplate.metadata?.labels,
-			...sandboxTemplateLabel(options.sandboxTemplateName),
-		},
-	}
 	return {
 		apiVersion: `${SANDBOX_API_GROUP}/${SANDBOX_API_VERSION}`,
 		kind: 'Sandbox',
-		metadata: { name: options.name, namespace: options.namespace },
+		metadata: {
+			name: options.name,
+			namespace: options.namespace,
+			...(options.annotations !== undefined ? { annotations: options.annotations } : {}),
+		},
 		spec: {
 			operatingMode: 'Running',
 			service: true,
@@ -706,11 +2037,87 @@ export function buildSandboxBody(options: SandboxBodyOptions): Record<string, un
 			...(options.template.volumeClaimTemplates !== undefined
 				? { volumeClaimTemplates: options.template.volumeClaimTemplates }
 				: {}),
-			podTemplate: { ...podTemplate, metadata, spec },
+			podTemplate: sandboxPodTemplate(
+				options.template,
+				options.sandboxTemplateName,
+				options.runtimeClassName,
+				options.podLabels,
+			),
 		},
 	}
 }
 
+/**
+ * The `spec.podTemplate` a directly created Sandbox carries: the template's,
+ * with this backend's overlays — the template label and whatever
+ * {@link composeAdditionalPodLabels} produced ({@link sandboxPodLabels}), and
+ * the configured `runtimeClassName`.
+ *
+ * Its own function because it is now built twice: once into the create POST
+ * by {@link buildSandboxBody}, and once into the JSON Patch that refreshes a
+ * standing workspace's pod template (`workspace.ts`). Two expressions of the
+ * same overlay would drift, and the one that drifted would report a workspace
+ * as off-template forever — the hash under
+ * `sandbox.namzu.ai/pod-template-hash` is taken over exactly this object, so
+ * a second spelling is a second revision.
+ *
+ * `podLabels` is on this signature rather than only on the create body for
+ * exactly that reason. A refresh rewrites `/spec/podTemplate` WHOLE, so a
+ * refresh built without them would PATCH the egress profile off a pod
+ * template that carries it — the replacement pod would come up selected by no
+ * per-profile policy, on a path where nothing re-checks the label, and the
+ * revision stamped beside it would be taken over a template the POST never
+ * writes, so `templateCurrent` would report drift forever.
+ */
+export function sandboxPodTemplate(
+	template: SandboxTemplateCopy,
+	sandboxTemplateName: string,
+	runtimeClassName?: string,
+	podLabels?: Readonly<Record<string, string>>,
+): SandboxPodTemplate {
+	const podTemplate = template.podTemplate
+	const spec =
+		runtimeClassName !== undefined
+			? { ...podTemplate.spec, runtimeClassName }
+			: { ...podTemplate.spec }
+	return {
+		...podTemplate,
+		metadata: {
+			...podTemplate.metadata,
+			labels: sandboxPodLabels(template, sandboxTemplateName, podLabels),
+		},
+		spec,
+	}
+}
+
+/**
+ * The labels a directly created Sandbox's pod will carry: whatever the
+ * template declares, plus this backend's own template label, which always
+ * wins because it is the label a policy selector is built to match.
+ *
+ * Its own function because the ingress check has to reason about EXACTLY the
+ * labels {@link buildSandboxBody} stamps, before the POST that stamps them.
+ * Two expressions of the same rule would be one rename away from a check that
+ * verifies a pod nobody creates.
+ *
+ * `extra` is `composeAdditionalPodLabels`'s map — the egress profile today.
+ * It is applied LAST, and so wins over both, for the same reason the template
+ * label wins over the copied template's own: it is a label the translated
+ * policy's selector is built to match, and a pod that matched the selector
+ * only sometimes would be a boundary that applied only sometimes.
+ */
+export function sandboxPodLabels(
+	template: SandboxTemplateCopy,
+	sandboxTemplateName: string,
+	extra?: Readonly<Record<string, string>>,
+): Readonly<Record<string, string>> {
+	return {
+		...template.podTemplate.metadata?.labels,
+		...sandboxTemplateLabel(sandboxTemplateName),
+		...extra,
+	}
+}
+
 /** The two halves of a `SandboxTemplate` a directly created Sandbox copies. */
 export interface SandboxTemplateCopy {
 	readonly podTemplate: SandboxPodTemplate
@@ -746,17 +2153,40 @@ export async function readSandboxTemplate(
 /**
  * Where the agent answers.
  *
- * Its own function, and the Service FQDN wins over a pod IP, because the
- * address outlives the pod: a suspended-then-resumed workspace comes back as a
- * new pod with a new IP behind the same name, and the transport re-resolves
- * the name on every dial. A literal IP baked into a long-lived handle is the
- * bug that would produce.
+ * Its own function, and under the default mode the Service FQDN wins over a
+ * pod IP, because that address outlives the pod: a suspended-then-resumed
+ * workspace comes back as a new pod with a new IP behind the same name, and
+ * the transport re-resolves the name on every dial. A literal IP baked into a
+ * long-lived handle is the bug that would produce — and it is exactly the bug
+ * `'pod-ip'` accepts, deliberately, in exchange for an address a host outside
+ * the cluster can resolve at all. That mode pays for it by re-reading the IP
+ * on every resume and once after a failed connect.
+ *
+ * `'pod-ip'` takes the address from `pod`, the record the bind token was just
+ * read out of, and NEVER falls back to `binding.podIPs`. The Sandbox's status
+ * is a second source that can name a different pod — the one a resume is
+ * replacing — and an address from one pod with a token from another is the
+ * mismatch that arrives as a flat `unauthorized`.
  */
 export function resolveAgentAddress(
 	binding: KubernetesSandboxBinding,
 	agentPort: number,
 	token: string,
+	options: {
+		readonly mode?: KubernetesAgentAddressMode
+		/** The live pod the token came from. Required by `'pod-ip'`. */
+		readonly podIP?: string
+	} = {},
 ): KubernetesAgentAddress {
+	if ((options.mode ?? 'service') === 'pod-ip') {
+		const podIP = options.podIP
+		if (podIP === undefined || podIP === '') {
+			throw new Error(
+				`kubernetes: sandbox ${binding.name} is configured with agentAddress: 'pod-ip', but the live pod its bind token was read from reported no status.podIP (nor a status.podIPs entry), so there is no address to dial. Every path that binds a pod POLLS for that address until the readiness deadline before this is reached — see readAddressedPod — so a pod that still reports none was never given one: a CNI that did not attach it, or a pod that never got past scheduling. Nothing is taken from the Sandbox's own status here on purpose — that IP may belong to a different pod than the token does.`,
+			)
+		}
+		return { kind: 'tcp', host: podIP, port: agentPort, token }
+	}
 	const host = binding.serviceFQDN ?? binding.podIPs?.[0]
 	if (host === undefined || host === '') {
 		throw new Error(
@@ -766,10 +2196,22 @@ export function resolveAgentAddress(
 	return { kind: 'tcp', host, port: agentPort, token }
 }
 
+/**
+ * Ready-or-not-yet, read off a `SandboxClaim`'s own status — plus the one
+ * "not yet" that is really a "no".
+ *
+ * The rejection check runs BEFORE the readiness check rather than after,
+ * because a refused claim is `Ready=False` forever and the readiness check
+ * cannot tell that from a cold start in progress. See
+ * {@link classifyClaimRejection}.
+ */
 function bindingFromClaim(
 	claim: SandboxClaimResource | undefined,
 	claimName: string,
+	metadata?: ClaimPodMetadataContext,
 ): KubernetesSandboxBinding | undefined {
+	const rejection = classifyClaimRejection(claim, claimName, metadata)
+	if (rejection !== undefined) throw rejection
 	if (!isConditionTrue(claim?.status?.conditions, READY_CONDITION)) return undefined
 	const bound = claim?.status?.sandbox
 	// Ready with no bound name is the controller contradicting itself; polling
@@ -804,38 +2246,141 @@ export function bindingFromSandbox(
 	}
 }
 
+/**
+ * The readiness poll ran out of budget — and nothing else. Every OTHER
+ * failure {@link pollForBinding} meets is rethrown as itself, so this class
+ * is an exact answer to "was it the clock?", which a caller that has to
+ * choose between two timeout messages needs and cannot get from the clock.
+ *
+ * Reading `remainingMs()` after the fact is NOT that answer: the expiry timer
+ * and `performance.now()` are different clocks, and a timer that fires a
+ * fraction of a millisecond early leaves a positive remainder behind an
+ * expiry that has already happened.
+ */
+export class ReadinessPollTimeout extends Error {
+	override readonly name = 'ReadinessPollTimeout'
+}
+
+/**
+ * What a failed readiness read is worth, decided by the caller.
+ *
+ * Optional, and absent means exactly the behaviour every caller had before:
+ * the first failure of any kind ends the poll. `workspace.ts` passes nothing
+ * and is unchanged; the acquire path passes
+ * {@link retryDelayForApiFailure} so one 429 on a shared cluster no longer
+ * fails a create that had fifty-nine seconds of budget left.
+ */
+export interface ReadinessPollBehaviour {
+	/**
+	 * Milliseconds to wait before reading again, or `undefined` to rethrow.
+	 *
+	 * The wait is spent on the SAME deadline as everything else in the poll,
+	 * so a retry consumes the readiness budget and can never extend it. A hook
+	 * that always returns a number therefore still terminates: the clock ends
+	 * the loop, not the hook.
+	 */
+	readonly retryDelayFor?: (err: unknown) => number | undefined
+}
+
 /**
  * Poll until `read` reports a binding. `read` returns `undefined` for "not
  * yet" and throws for a failure worth surfacing; the deadline owns every wait,
  * including the sleep between attempts, so an expired clock cannot be extended
  * by one more round trip. Shaped after ACI's `pollForRunningIp`.
+ *
+ * The only failure this raises on its own account is
+ * {@link ReadinessPollTimeout}; anything `read` throws travels out unchanged,
+ * unless `behaviour.retryDelayFor` claims it — in which case it is repeated
+ * inside the same budget and, if the budget then runs out, carried onto the
+ * timeout as its `cause`, so a poll that kept failing still says what it kept
+ * seeing.
  */
 export async function pollForBinding(
 	read: (signal: AbortSignal) => Promise<KubernetesSandboxBinding | undefined>,
 	deadline: OperationDeadline,
 	readiness: { readonly timeoutMs: number; readonly pollIntervalMs: number },
 	label: string,
+	behaviour: ReadinessPollBehaviour = {},
 ): Promise<KubernetesSandboxBinding> {
+	/**
+	 * The last failure that was retried rather than raised, and only while it
+	 * is still the truth: a read that succeeds clears it, so the timeout
+	 * carries a failure the poll was STILL meeting when the budget ran out
+	 * rather than a blip it recovered from twenty polls earlier. The
+	 * difference is not cosmetic — the acquire's reason is read off this
+	 * cause, so a stale one would report an API outage for a poll whose API
+	 * was answering fine.
+	 */
+	let lastRetried: unknown
 	while (deadline.remainingMs() > 0) {
+		/** Overrides the poll cadence for one round when the server named one. */
+		let nextDelayMs = readiness.pollIntervalMs
 		try {
 			const binding = await deadline.run(read)
+			lastRetried = undefined
 			if (binding) return binding
 		} catch (err) {
 			if (err instanceof OperationDeadlineExpired) break
-			throw err
+			const retryDelayMs = behaviour.retryDelayFor?.(err)
+			if (retryDelayMs === undefined) throw err
+			lastRetried = err
+			nextDelayMs = retryDelayMs
 		}
 		try {
-			await deadline.delay(readiness.pollIntervalMs)
+			await deadline.delay(nextDelayMs)
 		} catch (err) {
 			if (err instanceof OperationDeadlineExpired) break
 			throw err
 		}
 	}
-	throw new Error(`kubernetes: ${label} never became Ready (${readiness.timeoutMs}ms)`)
+	throw new ReadinessPollTimeout(
+		`kubernetes: ${label} never became Ready (${readiness.timeoutMs}ms)${
+			lastRetried !== undefined
+				? `; the last API failure retried inside that budget was: ${
+						lastRetried instanceof Error ? lastRetried.message : String(lastRetried)
+					}`
+				: ''
+		}`,
+		lastRetried !== undefined ? { cause: lastRetried } : undefined,
+	)
 }
 
 /**
- * The per-instance bind token: the backing pod's `metadata.uid`.
+ * The live pod behind a bound sandbox: its uid, and — read in the SAME
+ * answer — the IP it can be dialed at.
+ *
+ * One record rather than two reads because the two facts have to describe one
+ * pod. The uid is the agent's bind token and the IP is where that agent
+ * listens; taking them from separate GETs leaves a window in which a resume,
+ * an eviction or a node drain replaces the pod in between, and the handle
+ * then presents pod A's token at pod B's address. The guest answers that with
+ * a flat `unauthorized`, which says nothing about the race that caused it.
+ */
+export interface KubernetesBoundPod {
+	/** `metadata.uid` — the agent's bind token. */
+	readonly uid: string
+	/** `status.podIP`. Read by `agentAddress: 'pod-ip'`; absent is legal. */
+	readonly podIP?: string
+	/**
+	 * `metadata.labels` — what an ingress policy's `podSelector` actually
+	 * matches. Read off the SAME object the uid and the address come from,
+	 * for the same reason they are: a policy decision made about one pod and
+	 * a connection made to another is the mismatch this record exists to
+	 * prevent. The claim path reads it to ask which policies select the bound
+	 * pod — a directly created Sandbox's labels are known before its pod
+	 * exists — and BOTH paths read it to confirm the pod really carries the
+	 * labels this backend asked the controller for, which is the one thing
+	 * knowing them in advance cannot establish. See `ingress-policy.ts` and
+	 * `assertRequestedPodLabelsObserved`.
+	 */
+	readonly labels?: Readonly<Record<string, string>>
+}
+
+/**
+ * Find the pod a sandbox is currently backed by, and read both facts off it.
+ *
+ * `uid` is the per-instance agent bind token; `podIP` is where that agent
+ * listens, and only `agentAddress: 'pod-ip'` reads it.
  *
  * The pod is named after its Sandbox in agent-sandbox v1.0.2 — verified
  * against a running cluster — but that is an observation, not a documented
@@ -844,12 +2389,12 @@ export async function pollForBinding(
  * changing upstream is a second round trip through `status.selector`, which is
  * exactly what the controller publishes the selector for.
  */
-export async function readPodBindToken(
+export async function readBoundPod(
 	client: KubernetesClient,
 	namespace: string,
 	binding: KubernetesSandboxBinding,
 	signal?: AbortSignal,
-): Promise<string> {
+): Promise<KubernetesBoundPod> {
 	try {
 		const pod = await client.request<PodResource>(
 			'GET',
@@ -863,7 +2408,7 @@ export async function readPodBindToken(
 		// terminating this GET can answer with the pod that is leaving and a
 		// uid the new agent will refuse. On the acquire path nothing is
 		// terminating and the filter never fires.
-		if (uid && isPodLive(pod)) return uid
+		if (uid && isPodLive(pod)) return boundPod(uid, pod)
 	} catch (err) {
 		if (!(err instanceof KubernetesAlreadyGoneError)) throw err
 	}
@@ -879,7 +2424,7 @@ export async function readPodBindToken(
 		)
 		for (const pod of list?.items ?? []) {
 			const uid = pod.metadata?.uid
-			if (uid && isPodLive(pod)) return uid
+			if (uid && isPodLive(pod)) return boundPod(uid, pod)
 		}
 	}
 	throw new Error(
@@ -887,6 +2432,105 @@ export async function readPodBindToken(
 	)
 }
 
+function boundPod(uid: string, pod: PodResource): KubernetesBoundPod {
+	const podIP = readPodIP(pod)
+	const labels = pod.metadata?.labels
+	return {
+		uid,
+		...(podIP !== undefined ? { podIP } : {}),
+		...(labels !== undefined ? { labels } : {}),
+	}
+}
+
+/**
+ * {@link readBoundPod}, plus — under `'pod-ip'` only — the wait for an
+ * address to go with the token.
+ *
+ * Under the default mode this is the single read it has always been: one
+ * `GET`, in the same place in the same order, because a Service FQDN is
+ * published with the Sandbox and needs nothing from the pod but its uid.
+ *
+ * `'pod-ip'` has to wait, because a LIVE pod is not yet an ADDRESSED pod. A
+ * pod is created `Pending` and carries no `status.podIP` until the CNI has
+ * finished attaching it, and {@link isPodLive} accepts `Pending` on purpose —
+ * the resume path in `workspace.ts` binds its replacement pod long before that
+ * pod is Ready, because `Ready` stays True across the transition and the uid
+ * is the only transition signal there is. Refusing an address-less pod outright
+ * would therefore fail on the NORMAL path, in milliseconds, with the whole
+ * readiness budget unspent. So "live, no address yet" is polled on the same
+ * deadline as everything else on this path, and
+ * {@link resolveAgentAddress}'s own refusal is left as the post-deadline
+ * backstop for a pod that never gets an address at all.
+ *
+ * A failed READ stays fatal, exactly as it was: this is the acquire path,
+ * where nothing is being replaced and a pod that cannot be read is not a pod
+ * that is about to appear.
+ *
+ * `requiredLabels` is the second thing worth waiting for, and it is waited
+ * for in the SAME loop rather than in a second one: an egress profile is a
+ * label the CONTROLLER patches onto the pod it binds, so a pod read the
+ * instant it was bound can be live, addressed and not yet labelled. Two
+ * loops would be two deadlines and two answers to "is this pod ready to be
+ * admitted". Like the address, an expired clock hands the pod back as it is
+ * — the caller decides whether a missing label is fatal, and on the acquire
+ * path it is: see `assertRequestedPodLabelsObserved`.
+ */
+export async function readAddressedPod(
+	client: KubernetesClient,
+	namespace: string,
+	binding: KubernetesSandboxBinding,
+	deadline: OperationDeadline,
+	readiness: { readonly pollIntervalMs: number },
+	mode: KubernetesAgentAddressMode,
+	requiredLabels?: Readonly<Record<string, string>>,
+): Promise<KubernetesBoundPod> {
+	const required = Object.entries(requiredLabels ?? {})
+	const wanting = (pod: KubernetesBoundPod): boolean =>
+		(mode === 'pod-ip' && pod.podIP === undefined) ||
+		required.some(([key, value]) => pod.labels?.[key] !== value)
+
+	let pod = await deadline.run((signal) => readBoundPod(client, namespace, binding, signal))
+	while (wanting(pod) && deadline.remainingMs() > 0) {
+		try {
+			await deadline.delay(readiness.pollIntervalMs)
+			pod = await deadline.run((signal) => readBoundPod(client, namespace, binding, signal))
+		} catch (err) {
+			// An expired clock hands the incomplete pod back rather than
+			// replacing it with a bare "deadline expired": the caller's
+			// `resolveAgentAddress` (or `assertRequestedPodLabelsObserved`) then
+			// reports WHICH fact never arrived.
+			if (err instanceof OperationDeadlineExpired) break
+			throw err
+		}
+	}
+	return pod
+}
+
+/**
+ * The re-read a `'pod-ip'` handle follows a replaced pod with: one live-pod
+ * read, then the same address resolution acquire did.
+ *
+ * Built here rather than inside the transport because finding the pod is a
+ * CONTROL-plane act — the by-name GET, the selector fallback, the
+ * liveness filter — and the transport owns none of that. It is handed over as
+ * a closure so the transport can call it without learning what a Sandbox is.
+ */
+export function buildAgentAddressRefresh(
+	client: KubernetesClient,
+	namespace: string,
+	binding: KubernetesSandboxBinding,
+	agentPort: number,
+	mode: KubernetesAgentAddressMode,
+): (signal?: AbortSignal) => Promise<KubernetesAgentAddress> {
+	return async (signal) => {
+		const pod = await readBoundPod(client, namespace, binding, signal)
+		return resolveAgentAddress(binding, agentPort, pod.uid, {
+			mode,
+			...(pod.podIP !== undefined ? { podIP: pod.podIP } : {}),
+		})
+	}
+}
+
 async function readSandboxSelector(
 	client: KubernetesClient,
 	namespace: string,
@@ -939,17 +2583,31 @@ async function admitProbedSandbox(
 	config: KubernetesBackendInternalConfig,
 	options: SandboxBackendOptions,
 	probeTimeoutMs: number,
+	/**
+	 * Present exactly when `config.egress.perSandbox` is configured — which
+	 * is what makes `setNetworkPolicy` present on the handle. See
+	 * `per-sandbox-policy.ts`.
+	 */
+	setNetworkPolicy?: (policy: SandboxNetworkPolicy) => Promise<void>,
 ): Promise<Sandbox> {
 	const sandbox = buildKubernetesSandbox({
 		name: acquisition.binding.name,
 		rootDir: options.workingDirectory,
-		transport: new KubernetesAgentTransport(acquisition.agent),
+		transport: new KubernetesAgentTransport(acquisition.agent, {
+			// The backend opts in to the stream heartbeat; the transport
+			// option it sets stays undefined for every other tier.
+			heartbeatMs: resolveStreamHeartbeatMs(config.streamHeartbeatMs),
+			...(acquisition.refreshAgent !== undefined
+				? { refreshHandle: acquisition.refreshAgent }
+				: {}),
+		}),
 		release: acquisition.release,
 		renew: acquisition.renew,
 		ttlSeconds: acquisition.ttlSeconds,
 		...(config.onLeaseRenewalError !== undefined
 			? { onRenewalError: config.onLeaseRenewalError }
 			: {}),
+		...(setNetworkPolicy !== undefined ? { setNetworkPolicy } : {}),
 	})
 	try {
 		await probeSandboxPrivileges(sandbox, acquisition.binding.name, probeTimeoutMs, options.signal)