Repository navigation
PASS
Governs (arch 327e1b57. The checkout has no arch-map.md, so I found the sections by search, as the earlier rulings did): architecture.md Types in the identity plane, line 1291 (the client submits a digest-bound projection and the node independently recomputes it); Gatehouse §6.3 check 5, line 624 (a recomputed digest that differs is refused); the Box Record glossary, line 21 (the record keeps the projection digest).
Ruling: the PR adds one test, client_and_daemon_digests_match_vectors_and_mismatch_refused, in crates/minimald/src/box_manager/tests.rs, inside the task's one layer. It changes no production code, so rules 7, 8 and 10 are not engaged. The client side follows the same of_spec → project().digest() path as stamp_client_projection_digest (crates/minimal/src/cmd/boxes/create.rs:3266-3273). The daemon side is the existing BOX-071/BOX-072 gate in crates/minimald/src/box_manager.rs, which refuses before the record exists and logs client_projection_digest/daemon_projection_digest. The test asserts the stored attrs, the refusal type, the unchanged box count and the log fields. It proves the cited rule without relaxing it (rule 3). This is consistent with #3442 and #3441.
Notes: the test adds v = BOX_SPEC_VERSION to the wire spec, because the golden expanded specs carry no v. That matches what min sends today, and v is outside the projection. The author names #2410 as the task that must add v when it sends the expanded spec. That is work for #2410, not a residual of this PR. The client path is a copy in the test, because minimald does not depend on the minimal crate. A drift between the copy and stamp_client_projection_digest is caught only by the golden vectors on each side.
Not checked: I did not compile or run the test (out of bounds). The PR reports the named test at exit 0. The full agent-verify run timed out on env::tests::minenv_run_forwards_stdin_to_the_task and cancelled on an unnamed failure. That test is in another module, and an added test cannot change it. The PR is a draft for that reason, and the merge gate is CI, not this ruling.
Ruled at 85ffbf1 against gominimal/arch@327e1b57d16b by claude-opus-5-5; arch_rulings 1322, job 621e5eff-353f-4580-af13-2eb0439943a6.