Repository navigation
FAIL
Inputs. Re-review of #3619 at 8b16de35, on arch 4f2eb87d. The three earlier rulings (31c6b775 twice, 7e880675) were all FAIL on arch 8b824707, so since-conditions.json is empty and no condition is open. The why of those rulings is cut short in the ledger, so I read pr.diff whole and ruled on the full change again. since.diff starts at 7e880675. It holds two merges of main and five PR commits. e4ad786c reverts the earlier network table refusal. 9d71f7c1, d8e86338 and 3d2fc561 add merge_network and the per-key origins. ea7f4e0f amends BOX-048 and the NET S9a/AC1 note. No findings.json or arch-map.md is staged, so I found the sections by search. The task marker (D3597) has no covers: line, so I ruled on its three layers, as the first ruling did.
Governs. architecture.md minimal.toml layout. Line 1264 (Sections) says [defaults] merges into every entry, lists union, [env_vars] merges by name, and other scalars and tables are overridden. Line 1294 (Expansion) says a table is replaced wholesale, not merged key by key. It uses network as its example: an entry that sets network.egress drops an earlier mode and must restate it. The same line names [env_vars] as the one table exception. The create-time network flags are the one exception for network, and they are not layers of minimal.toml. example-minimal.toml lines 25-26 repeat the rule.
Ruling. crates/mfile/src/expand.rs merge_network merges network by key across layers, so [defaults.session.network.bep] keeps [defaults.network.egress]. That is the behaviour line 1294 rules out. It also adds a rule the design does not state: a later mode = "none" clears an inherited egress and ingress. The CLI layer (box_spec.rs, create.rs ProjectEgress.mode_layer, task.rs) reports per-key origins that exist only because of that merge. The docs layer amends BOX-048 and the spec 29 decision text to say network merges by key, and cites gominimal/arch#188 as the decision. The design of record at 4f2eb87d holds no such decision. No commit, branch or file in the checkout names #188, and line 1294 still states the opposite. Rule 9 applies: the design is the record, and a spec line that contradicts it does not make the code conform. No condition inside this task can make the change conform, because the merge is the whole task. So the ruling is FAIL, the same as the three earlier rulings on this PR.
Path forward. If gominimal/arch#188 decided that network merges by key, land that decision in architecture.md lines 1264, 1288 and 1294 and in example-minimal.toml lines 25-26. Then push again, and the next ruling reads the new arch commit. The design text must also state the mode = "none" clearing rule, or the code must drop it. That is a design question for #188, not for this PR.
Not checked. Whether gominimal/arch#188 exists or what it says. The checkout does not carry it, and I cannot read GitHub from here. I did not build or run tests. The ruling does not need a build: the deviation is in the merge rule, not in whether the code compiles.
Ruled at 8b16de3 against gominimal/arch@4f2eb87d5d12 by claude-opus-5-5; arch_rulings 1570, job d947c510-281d-48fd-b2f2-770582d6b4d1.