Repository navigation
PASS (the design is silent)
Inputs. Re-review of #3702 at 9ef5b5aa, on arch 4f2eb87d. That is the same arch commit as the four earlier rulings on this PR. since.diff starts at 23727ba6, which was a PASS. since-conditions.json is empty. That PASS closed any conditions, so no condition is open and the compare is enough. No findings.json or arch-map.md is staged, so I found the sections by search.
Compare. One commit (9ef5b5aa). In scripts/install-host-classifier.sh ct_mark_rule_owners(), the line for a rule that a well-known mark range accounts for no longer says that the rule's packet-mark bits are the ones "which Tailscale marks connections with". It now says that the program that uses the range can load the rule, and that another program that loads it must write only the connection-mark bits it uses. The three tests in crates/minimald/src/net/classifier.rs that pin this text follow the new wording. Both the positive and the negative assertions changed (install_refuses_tailscale_nft_mode_connmark_with_its_match and the two neighbouring tests for a wide mask and a one-bit mask).
Governs. The design names the cgroup classifier as one netfilter cgroup-match ruleset (specs/networking/deployment-and-egress-gateway.md line 54). It says nothing about connection-mark conflicts with other programs, the install-time scan, or the text of the refusal. A search of architecture.md and specs/networking/ for connection mark, ct mark, connmark and mark range finds no line. So this is a PASS by absence.
Ruling. The change is wording only. Rule 7 (fail closed): the refusal still ends the install through die in scan_ct_mark_conflicts, and I confirmed that at head (line 1809). Rule 8 (placement): the edit stays inside ct_mark_rule_owners(), where the earlier heads put the owner lines. The new text claims less than the old text, because a mask match does not prove which program loads the rule. That makes the audit line more exact, not less. The negative assertions still check for the range sentence, so the tests do not pass by checking for nothing (rule 3).
Not checked. I did not run the tests, by the no-build rule. The root-integration test install_refuses_a_whole_mark_ct_writer is ignored in the default run, and the PR body says it was not run. The user-facing wording follows AGENTS.md (User-facing text), which is a repository rule and not a design line, so it does not decide this ruling.
Ruled at 9ef5b5a against gominimal/arch@4f2eb87d5d12 by claude-opus-5-5; arch_rulings 1580, job 87b7e91a-6d09-49d0-b1e1-ba8f9705591b.