docs(handoff): record the W3 experiments and decision packet

[novelKR]

Summary

This is the non-normative W3 record for CS-DG-UPSTREAM-ADAPTER-WORK-SPEC-1 v1.0. It records:

• the owner-approved merges (docs(handoff): record the CS-RG boundary revalidation #13, docs(handoff): record the upstream-adapter W0-W2 packet #16, docs(notes): point to the DevGuard upstream-adapter packet CodeSpace#78) and the evidence housekeeping;
• bounded experiments A, B and C (sections 13.1–13.3 of the W0–W2 packet), with their frozen protocols, amendments,
  and failed or invalid attempts;
• the updated candidate status;
• the owner decisions now requested.

It stops at the next owner decision. Nothing here approves an architecture, dependency, pin, contract,
implementation, upstream submission or merge.

What changed

• docs/handoff/2026-09-28-w3-decision-packet.md (new).

Key results (macOS, one host; behavioural evidence, not proof):

• A. A permit-preserving preparation, consumed by Codex's unmodified PTY ChildFds::Attached with DevGuard's
  unchanged helper, wire and authority, preserved single use, first-claim-wins, fencing, no replay and the
  known-not-started rules. The binding needed no PID accessor.
  • The frozen FIFO carriers do not work with the unchanged poll-based readers on this host. The amended run used
    today's carriers and an atomic regular-file permit.
  • With Attached, a PTY signal death still reports exit 1.
• B. A generic scratch patch to codex-utils-pty (not submitted) keeps the backend as the only reaper. It
  observes the exit before reaping, bounded by a fixed timeout.
  • It turned the reap-first survivor case into normal release: 50 of 50 healthy runs returned their reservations.
  • Stalls, crashes and an unavailable authority degraded to the existing conservative states.
  • On the PTY path, a same-group background descendant receives SIGHUP when the root exits.
• C. Unrelated std and Tokio children inherited DevGuard's session socket and today's launch carriers while they
  were being created. A holder of an inherited session could read the permit-bearing reply and inject requests.
  • Kernel close-by-default in the spawner removed every leak class in both samples.
  • An atomic file permit removed the permit class.
• Also: the CodeSpace descriptor-hygiene issue Investigate descriptor inheritance between concurrent spawns on macOS CodeSpace#79 was opened. It is non-implementation and
  justified on CodeSpace's own terms.

What did not change

• No contract, design document, AGENTS.md, NOTICE, validation script, manifest, lock or Codex pin.
• No CodeSpace code or policy page. The CS-RG hold stays in force.
• No service, credential, journal or host setting: the installed LaunchAgent kept the same pid throughout.
• Nothing was submitted upstream.
• The experimental DevGuard branch stays local and was never pushed.

Verification

• python3 scripts/check_docs.py passed (17 pairs, 48 units, 25 groups). test_check_docs.py passed.

• git diff --check is clean. The new file contains no local paths, and every table row has the header's cell
  count.

• An independent read-only audit checked 204 claims against the sealed evidence, git and GitHub. It found:

  • two framing errors: amendment-changed cases presented as frozen passes;
  • sixteen minor points.

  All are fixed in this head. It confirmed every evidence digest, protocol digest and merge fact.

• scripts/validate.py was not run locally; CI runs it.

Merge and rollback

• Merge only at the exact head the owner approves (packet section 6, decision 1). Merging records the packet; it
  approves nothing proposed in it.
• The change is documentation only, so rollback is a revert of the merge commit.

Handoff

• Start here. Read DevGuard #14's latest session-close comment. Then read this PR at its exact head 70c814c93a96f383e3241b6817fe6d42aea9ea48.
• CodeSpace side. Read the latest session-close comment on novelKR/CodeSpace#76 and Investigate descriptor inheritance between concurrent spawns on macOS CodeSpace#79.
• Status. This PR is non-normative and unmerged. Merging it is an owner decision, and nothing in this handoff implies it.
• Final CI at session close (not rerun):
  • Push run 36408700221 on 70c814c succeeded on macOS 14 and Ubuntu 24.04.
  • Pull-request run 36408760157 succeeded on Ubuntu 24.04 and failed on macOS 14, in a_drain_that_does_not_finish_in_time_keeps_the_current_release_and_its_charges (crates/daemon/tests/upgrade.rs:372). The failing assertion is started.elapsed() < Duration::from_secs(10), after the expected ResourceUnavailable and "did not finish" result.
  • Both runs tested the same repository tree, and that tree passed the test on macOS in the push run.
  • Handoff: CS-RG hold, boundary revalidation and pending owner approvals (2026-09-27) #14 classifies this narrowly, as a timing-sensitive failure not tied to this documentation-only change.
• Decisions. The unresolved decisions are in section 6 of the packet. Their current interpretation, and two new follow-up tracks that have not started, are on Handoff: CS-RG hold, boundary revalidation and pending owner approvals (2026-09-27) #14.
• Stop boundary. Until the owner decides, there is no merge, experiment, implementation, pin or dependency change, normative PR, upstream submission, or service or credential change.