Repository navigation
FAIL
Inputs. First ruling on #3619 at 31c6b775, on arch 8b824707 (the merge of gominimal/arch#177). No since.diff, since-conditions.json, findings.json or arch-map.md is staged, so I found the sections by search. The task marker (D3597, origin bug, fixes #3600) has no covers: line. I split the task into its three layers (expand, cli, docs) and ruled on each.
Governs: architecture.md minimal.toml layout. Line 1223: [defaults] merges lists by union and [env_vars] by name, and other tables are overridden. Line 1253 (Expansion): "a table is replaced wholesale rather than merged key-by-key", with network = { egress = … } dropping an earlier mode as the example. The only exceptions it names are the create-time network flags (arch#166) and [env_vars] (arch#177). Policy layers and precedence, line 1297, says the same. example-minimal.toml lines 88-89 show an entry that restates mode for this reason.
Ruling. merge_network in crates/mfile/src/expand.rs makes network merge by key, which is the key-by-key merge that line 1253 excludes, for its own example. The CLI attribution edits (entry_egress, ProjectEgress.layer, setting_table, network_sources) only report that merge. The spec edits (BOX-048 and the spec 29 decision at line 830, and the spec 18 NET annotation at line 659) move the spec away from the design. Rubric rule 9 forbids that. The design is not ambiguous and two sections do not conflict. The design rules out the behaviour that #3600 asks for. No condition on this PR can make a sub-table merge conform, so the verdict is a deviation. The path forward is the escalation: an owner ruling in gominimal/arch that adds network as a merge exception, as arch#166 and arch#177 did for the flags and [env_vars]. With that ruling merged, this change can come back for review against the new arch commit. If the owner keeps the wholesale replace, #3600 closes as working as designed.
Precedent. #3632 (ab04767c): the vars merge exception reached spec 29 after arch#177 put it in the design. This ruling follows that order. No ruling in the ledger accepts a merge exception that the design does not state.
Other notes for a later round. The module doc at crates/mfile/src/expand.rs lines 7-9 and the BOX-048 property: line (spec 29 line 288) still state the wholesale replace, so the head disagrees with itself. A sub-table merge lets a Box Type's default mode survive an entry that sets only egress. That changes the network mode in the Box Spec projection. The design ruling must state that effect.
Not checked. I did not build. I did not check whether the node-side recompute of the projection (BOX-072) uses the same mfile fold, or whether a version skew between client and node expanders with different merge rules gives a digest mismatch. The PR's agent-verify run exited 100 on otel_cli exec_run_attach_and_task_run_share_the_commands_trace. That failure does not bear on this ruling.
Needs a design decision
- #3600 asks that a
[defaults.session.network.bep]not drop a[defaults.network.egress]. architecture.md line 1253 (minimal.toml layout, Expansion) says a table is replaced wholesale, and usesnetworkas its example: a layer that sets onlyegressdrops the earliermode. example-minimal.toml lines 88-89 say the same. Isnetworkto merge by sub-table (each key belownetwork, such asmode,egress,bep,ingress, from the most specific layer that sets it, withegressand the flategress_*keys as one key), as[env_vars]merges by name? If yes, the design must also say: whethermodefrom a Box Type's defaults now survives an entry that sets onlyegress(this changes the projected network mode), howmin box specnames the layer of each key, and whether the rule applies to the loadout layer. If no, #3600 is not a defect, and the issue closes with the documented workaround: restate the earlier sub-tables in the later layer.
Ruled at 31c6b77 against gominimal/arch@8b824707cde6 by claude-opus-5-5; arch_rulings 1442, job 932e3581-ae03-4a1e-9105-92f7ddc213d8.