ADR-101 — Forward-only scrub of host and user names + guard classes

• Status: Accepted on merge (security / compliance guardrail, sponsor decision 2026-10-13 20:49 CEST, relayed by DSO). Not a kernel change.
• Date: 2026-10-13
• Follows: ADR-093 (cloud identifiers + scripts/check-public-ids.sh). ADR-099 is taken by the open draft PR #152; this ADR takes 100.
• No history rewrite (sponsor: forward-only). The replaced names remain in earlier commits and merged PR diffs.
• _start (0x4008_0000) untouched. No kernel logic change: src/ and linker.ld change in comments only.

Context

ADR-093 scrubbed cloud identifiers and the local OS user name in host paths. It did not classify machine names (the nicknames of the two developer workstations, used in docs as if they were roles), host names that tools printed into committed logs (an agent VM host name, an EC2 private host name, GitHub-hosted runner host names inside uname -a lines) or a machine ID (a UUID, in full once and as an 8-hex prefix four times). This repository is public. Two scripts would also print a host name into future logs: scripts/b2-metal-userdata.sh (uname -a) and scripts/qemu-nmi-probe.py (socket.gethostname()).

Decision

1. Replace with role names or placeholders (values are not listed here):
   • the amd64 developer workstation → “amd64 workstation”; the arm64 developer workstation → “arm64 workstation” (with “the” where the grammar needs it);
   • agent VM host name → <agent-vm-hostname>; EC2 private host name → <ec2-private-hostname>; GitHub runner host names → <gha-runner-hostname>; machine ID → <machine-id>. “agent-box” stays: it is a role name, not a host name.
2. Stop new leaks at the source: b2-metal-userdata.sh prints uname -srm (as ADR-095 already did by hand); qemu-nmi-probe.py prints host-os=<sysname> <machine> instead of the host name.
3. Guard classes (fail closed, CI-gated through scripts/qemu-smoke.sh as before). scripts/check-public-ids.sh gains eight classes: vm-hostname, ec2-private-hostname, uname-hostname (a host-name-shaped token between Linux/Darwin and a version), gha-runner-hostname, user-home-path (C:\Users\<name>, /Users/<name>, /home/<name>; /home/runner and <placeholder> allowed), wsl-user (wsl … -u <name>; root and <placeholder> allowed), uuid, and machine-name. Machine names are matched by SHA-256 of the lower-cased token, so the guard does not spell them out. That keeps them out of plain-text search; a hash of a short name can be brute-forced, so it is not secrecy. On a hit the guard still prints only path:line class. --self-test now plants a synthetic token for all 15 classes (including a runtime-hashed fake machine name in two cases) plus a line of allowed forms that must not be flagged; marker public-ids: self-test ok classes=15.

Counts (by class, no values)

Guard of this branch run on unmodified main af5a9d1 (distinct path:line class pairs) and a plain grep for occurrences:

New guard on main: FAIL hits=151. On this branch: public-ids: ok … hits=0 (bash and dash).

Limits (non-claims)

• Shapes and a short hashed name list, not DLP. A new machine name needs its hash added. Short role-like words are not flagged.
• Git history, PR descriptions, issue comments and CI logs are not scanned or changed.
• Git author metadata and the GitHub owner name in repo URLs stay (public and intended).
• Not a “secure” or “compliant” claim.