[Enrique Saurez (esaurez)]

Refs #1116

Hand the real layered init's retained root, control endpoint, CLI identity,
account defaults, mapper, and optional source root to the existing managed
dispatcher and runtime. Retain init-owned helper failure and shutdown behavior,
prelaunch ownership, cancellation, and the legacy sandbox path.

All twelve qualified SL7.6b commits were replayed exactly once directly onto
#1397 at 29fca59317ffb67d8a3c89061faddf224c086051, without conflicts,
ancestor commits, or merges. Original messages, trailers, authors, patch
operations, modes and unsigned signature states are preserved. Their only
range-diff difference is unchanged allowlist hunk context from the new parent.
The OpenVMM pin remains 14dc4116.

The independent review confirmed Copilot's blocking-output finding. The coupled
fix makes only runtime-backed control descriptors nonblocking and polls writable
readiness alongside the same Runtime's signal descriptor. It retains statuses
collected during writes, hands them to the execution driver before each event,
and preserves orderly termination during ATTACH and RESET acknowledgement.
Legacy sessions remain unchanged; no timer-retry workaround, second reaper,
public API or wire-format change was introduced.

The first fix commit accidentally placed the new poll regression inside an
assertion's token tree. The independent review caught this; the append-only
correction 9445f16b places it at module scope without changing its assertions.
Actual Linux compilation then caught a slice/array mismatch in the new Runtime
test's frame decoder. c7e97160 uses the existing session-test conversion to the
decoder's fixed-size header type. Production and framing assertions are unchanged.
Both introduced test defects and their superseded results remain historical.

The full actual-base delta is 313 CORE, seven counted paths, fifteen raw paths:
twelve preserved commits plus three coupled fix commits. The user approved the
existing poll helper/test extension and necessary additional lines for this task.
All JSON and correction lines are counted, without extraction/category discounts.

Existing tests

The full structured comparison flags eight existing assets, relocates three
unchanged, adds 52 assets and extends one. The original seven flagged assets remain:

• init::tests::{init,serve} keep the old helper/signal scripts in test-only
  drivers of production preparation, collection and shutdown.
• Harness::spawn_at delegates to a shared constructor for the same dispatcher.
• Harness::read_outer delegates to the same extracted frame reader.
• isolate_process_test includes ignored tests only in its exact named child
  invocation, so privileged regression cases are actually runnable.
• The existing managed lifecycle test additionally checks real layered init with
  zero/two mappings, identity/defaults, reconnects and cancellation; its original
  workload/shutdown checks remain.
• The obsolete Runtime::wake hook disappears with the idle loop; the existing
  init hook allowance tightens from eight to seven.

The eighth flagged asset is the existing poll-order test: its expected readable
event also asserts writable == false. New focused regressions cover writable
readiness/backpressure, SIGTERM and helper-failure precedence while output is
blocked, pending termination before ATTACH, and correct delivery of a status
collected during a write. No assertion, deadline or existing conditional skip was
relaxed. The new Runtime cases retain their real CAP_SYS_NICE/CAP_SYS_RESOURCE
prerequisite.

Validation

Current head: c7e971607a0ea3b638800b4d90a3fd549685c42b.

• Full current-head CI run 42697389932, attempt 1: success. Current Copilot
  review 5832613919: COMMENTED, with zero unresolved review threads at the
  final inspection. No formal approval is claimed.
• Independent whole-owned source review: CLEAR, actual Opus 5.5/MAX/LONG.
  The user approved the retained reviewer's audited read-only path; actual calls
  and source/native/index immutability were checked. There was no technical
  tool-deny fence, and this source-only review did not substitute for execution.
• Windows: all 29 Rust command rows passed in a fresh whole confirmation with
  strict guard exit0 and no owned descendants. Standalone Python/quality passed:
  1823 unit tests, 70 conditional skips, plus applicable lint/type/format,
  layout, kernel-inventory and PowerShell checks, strict guard exit0.
• Linux: format, five clippy variants and the focused poll regressions passed.
  The full all-feature library run passed 2091 tests, 51 ignored, with
  default test parallelism. All 33 remaining Rust command rows passed, including
  applicable MSRV/cross-target, public API, guest and adapter matrices.
  CLI/quality passed 1823 unit tests, 30 conditional skips, Specula's
  25 tests, and applicable shell/type/lint/layout/kernel/PowerShell checks.
• All ten selected privileged Runtime cases actually executed and passed,
  including the four new regressions: one passed, zero failed, zero ignored
  per invocation. These were current frozen-binary runs under narrowly approved
  private-namespace privileges, not early-return skips or old-head evidence.
• Fresh current Alpine/Ubuntu guests, distro, real layered-init fixture and
  runtime-adapter guest were assembled and input/provenance checked. Native
  OpenVMM/kernel inputs were reused only after matching provenance validation.
  Eighteen host runners per platform were prebuilt and hash-bound; hardware
  execution used a strict compiler-free bridge, with no own compiler/VM overlap.

Fresh hardware acceptance completed on Windows/WHP (user-declared bare metal)
and Pedro/MSHV (configured virtual machine):

Acceptance	WHP	MSHV
Real layered init, maps0 and maps2	Each 1 passed, 0 ignored	Each 1 passed, 0 ignored
Whole default microVM suite	57 passed records, 51 distinct scenarios	57 passed records, 51 distinct scenarios
Public execution/lifecycle/proxy/RAM-upper checks	27 / 41 / 12 / 14	27 / 41 / 12 / 14
Canonical runtime-adapter suite	156 cases, 11 targets passed	156 cases, 11 targets passed

Both backends ended with no owned VMMs. MSHV VMMs ran after dropping to UID1001,
not as root. The full MSHV operation and both WHP layered cases passed their
unchanged resource guards. The user explicitly authorized bypassing resource
guard/admission for the remaining WHP default and canonical tests. Those are
actual test passes under that exception, not resource-guard passes. Original
test assertions/deadlines, bounded execution, source/artifact checks, owned Job
containment and teardown checks remained in force; no global guard was changed.

Historical failures remain recorded, not waived or relabelled:

• Current-head initial GNU runs exposed long Unix-socket paths (77 failures),
  then an unmapped inherited process group (one failure). User-approved private
  /tmp binding and an own session/group before Bash PID1 corrected only the QA
  environment; the unchanged-source, default-parallel whole run then passed.
• An earlier current-head Windows Rust attempt completed command rows but failed
  its guard with exit125. The separate whole confirmation passed the unchanged
  guard; the earlier cause remains unproven. Superseded 9445f16b also retains
  its E0308 test-helper compilation failure, Windows guard125 and telemetry
  timeout; none qualifies this head.
• First candidate 9698af55 passed its CI but retained the actionable blocking
  write finding. Its first Windows Python run had 63 failures, 29 errors and 71
  skips, including DLL-initialization errors; their cause remains unproven.
  The superseded 17ca29e3 interruption and the original source's ETXTBSY
  history, conditional skips and limitations remain historical.

A separate red-baseline executable was not built or run. No performance
comparison is inferred from these acceptance results.

This PR remains Draft. Its current head received exactly one authorized
temporary Ready/full-CI/Draft cycle; no additional CI rerun was used for local
qualification. Final Ready and merge remain the user's decision.

[github-actions]

This pull request changes existing tests

It modifies or removes 12 existing test asset(s), listed below. Review each with care: a change to a test can hide a regression that the test caught. See doc/contribute.md#changes-to-existing-tests.

Test code

Change	What	Before	After
modified helper	fn init in guest::init::tests	nvx-rs/src/guest/init/tests.rs:26	nvx-rs/src/guest/init/tests.rs:26
modified helper	fn serve in guest::init::tests	nvx-rs/src/guest/init/tests.rs:31	nvx-rs/src/guest/init/tests.rs:31
modified helper	fn spawn_at in guest::managed_agent::dispatch::tests, impl Harness``	nvx-rs/src/guest/managed_agent/dispatch/tests.rs:70	nvx-rs/src/guest/managed_agent/dispatch/tests.rs:70
modified helper	fn read_outer in guest::managed_agent::dispatch::tests, impl Harness``	nvx-rs/src/guest/managed_agent/dispatch/tests.rs:210	nvx-rs/src/guest/managed_agent/dispatch/tests.rs:214
modified helper	fn isolate_process_test in guest::sys::process::tests	nvx-rs/src/guest/sys/process/tests.rs:13	nvx-rs/src/guest/sys/process/tests.rs:13
modified helper	fn reach_guest in vm_clone (tests)	tests/runtime-adapter/tests/vm_clone.rs:4788	tests/runtime-adapter/tests/vm_clone.rs:4788
modified helper	fn reach_guest in vm_layered (tests)	tests/runtime-adapter/tests/vm_layered.rs:3019	tests/runtime-adapter/tests/vm_layered.rs:3017
modified imports	the imports of vm_clone (tests)	tests/runtime-adapter/tests/vm_clone.rs:206	tests/runtime-adapter/tests/vm_clone.rs:206
modified imports	the imports of vm_layered (tests)	tests/runtime-adapter/tests/vm_layered.rs:41	tests/runtime-adapter/tests/vm_layered.rs:41
modified test	fn poll_reports_input_in_the_callers_order in guest::sys::poll::tests	nvx-rs/src/guest/sys/poll/tests.rs:7	nvx-rs/src/guest/sys/poll/tests.rs:7
modified test	fn a_managed_guest_runs_request_scoped_workloads_until_it_stops in managed_lifecycle (tests)	nvx-rs/tests/managed_lifecycle.rs:369	nvx-rs/tests/managed_lifecycle.rs:494
removed hook	the test hooks of fn wake in guest::init (impl Runtime)	nvx-rs/src/guest/init.rs:860

Relocated unchanged: 3; added: 60. Compared the merge base 29fca59317ff with 061b99e153b5.

[Copilot]

🟡 Changes recommended

A critical blocking control-output write can leave PID 1 unable to handle shutdown or helper failure when the host stops draining output, requiring nonblocking writes and backpressure regression coverage.

1 open issue

• Blocking control writes can prevent PID 1 from handling signals · New

What changed in this PR

This PR advances #1116 by routing nvx-init’s prepared layered roots through the managed agent’s existing control dispatcher, sharing execution and mapping handling with blockless VMs while retaining init-owned shutdown and helper supervision.

Changes:

• Pass retained mapping handles, CLI identity, and account defaults into the managed dispatcher.
• Integrate init’s status collector and signal handling into control reads and workload execution.
• Add layered-runtime regressions and lifecycle checks, while moving the obsolete idle-loop driver into test-only code.

File	Description
scripts/​test-layout-allowlist.json	Tightens init’s test-hook allowance.
nvx-rs/​tests/​managed_lifecycle.rs	Adds real layered-init mapping, identity, reconnect, and cancellation checks.
nvx-rs/​src/​guest/​sys/​process/​tests.rs	Allows explicitly selected ignored tests in isolated children.
nvx-rs/​src/​guest/​managed_agent/​tests.rs	Tests layered CLI identity parsing.
nvx-rs/​src/​guest/​managed_agent/​session.rs	Integrates init supervision and status retention into control reads.
nvx-rs/​src/​guest/​managed_agent/​mappings/​tests.rs	Tests retained-export requirements.
nvx-rs/​src/​guest/​managed_agent/​mappings.rs	Adopts prepared layered mapping state and handles.
nvx-rs/​src/​guest/​managed_agent/​exec.rs	Routes workload collection through the shared session.
nvx-rs/​src/​guest/​managed_agent/​dispatch/​tests/​runtime.rs	Adds privileged shutdown, helper-failure, and status-ownership regressions.
nvx-rs/​src/​guest/​managed_agent/​dispatch/​tests.rs	Shares harness helpers and tests layered dispatch.
nvx-rs/​src/​guest/​managed_agent/​dispatch.rs	Adds layered configuration and runtime-backed serving.
nvx-rs/​src/​guest/​init/​tests.rs	Keeps idle-loop test drivers outside production code.
nvx-rs/​src/​guest/​init.rs	Replaces the production idle loop with managed dispatch.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

The new test function is declared inside the expected array passed to assert_eq!, preventing Linux guest unit tests from compiling.

1 open issue

• Move backpressure test out of the expected array · New

1 resolved since last review

• Blocking control writes can prevent PID 1 from handling signals

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Changes to PID-1 control, reaping, and shutdown across switch-root require privileged regression tests and real layered-init boot validation that remain pending and cannot be established by source inspection alone.

No open issues 🎉

1 resolved since last review

• Move backpressure test out of the expected array

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Production PID-1 control, child-status collection, and shutdown changes still await current-head privileged Runtime regressions and unhooked layered-init lifecycle validation on qualified hosts.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Human assessment is needed because caching a ptrace stop can cause the leader’s terminal status to be discarded, producing a lost-status error instead of the actual termination result.

No open issues 🎉

_{Review effort: Balanced}