A05: minimal bolt groups joining two parts - #109

Verdict: approved

A05 bolt groups promotion, nine commits, net diff reviewed against upstream/master.

• Model admission (crates/model/src/command.rs): complete and fail-closed — label, two distinct parts (DV-REL-006/007), the group is not its own part, finite diameter in (0, 1000], count 1..512 with the right code split (InvalidArgument for 0, LimitExceeded above), finite pitch under the dimension cap, origin under the coordinate cap, and a unit direction by squared-norm tolerance. The fingerprint is versioned (BeamCraft/A05/CreateBolt/v1), length-prefixed and canonical-f64 encoded.
• Engine decision (crates/engine/src/bolt.rs): pure and bounded over the weld Parts index; every bolt position origin + direction * (pitch * i) must lie inside both parts and a supported part without bounds refuses (None => false), never guesses; NaN bounds fail the comparisons and refuse. The whole group is refused together, and the tests cover missing/wrong-kind/taken identities, the 13th bolt leaving the plate, the out-of-plane origin and the count-1 outside origin.
• Connection reading (apps/beamcraft-cli/src/weld.rs): a connection with any steel:bolt: property (whole prefix, not a loose substring) is a bolt group and never a weld; a connection with neither set is refused by the validator before the walk. kinds was widened to pub(crate) for the bolt adapter, nothing else moved.
• The converge commit is the right surface discipline: the verbatim cherry-picks had carried fork-only neighbours (bom module registrations, A09 mark constants and their tests, the fork dispatch arm); the branch now registers and dispatches bolt alone, after the weld arm.
• The mapping-pin chain is correctly re-based on the upstream descriptor: pre_a05_descriptor() is the current descriptor minus the bolt-properties line with the connection phrase reverted, the placeholder was replaced by the recomputed upstream pre-A05 digest (the fork value covered a mark line upstream never had), and the D1/D2 pin and the all-pins-still-accepted test chain through it consistently.
• limits.rs was rewritten to LF: upstream had mixed CRLF/LF there, every other touched file was already LF, so this normalises the one outlier; the real change is the two bolt caps.

No CI checks have reported on the branch yet; this approval rests on the review above and on the same feature having run in the fork since its A05 slice.

Verdict: changes requested (one rebase-introduced pin-chain regression)

The A05 code itself is unchanged from the version I approved; what broke is the mapping-pin chain, which was rebased from the fork history onto a master that already carries the native A09 (#103, merged before #105).

On this branch SCHEMA_DESCRIPTOR = A07.2-era descriptor + native A09 line + A05 bolt line. pre_a05_descriptor() chains through pre_a09_descriptor() (current minus the A09 line), giving:

• accepted: current; pre_a09 = descriptor-with-A05-without-A09 — a mapping no released master ever shipped; pre_a05 = the A07.2-era fdaf890...; and the older pins;
• not accepted: the descriptor of the current released master (fdaf890 + the A09 numbering line) — which is exactly the digest every store created upstream since #103 carries.

pin_accepted is load-bearing on the edit path (authored/edit.rs refuses a layer whose saved bimSchemaDigest is not accepted with UnsupportedVersion), so an A05 build refuses to edit any store authored by current upstream master. The test stays green because it pins the same wrong constant (fdaf890, now described as "pre-A05" and duplicated in the all-pins list as "pre-A09"), and the comment "A05 predates A09" describes the fork history, not this one.

The fix is small: define the pre-A05 descriptor as SCHEMA_DESCRIPTOR minus the two A05 edits (that is the live master descriptor and the real pre-A05 mapping), accept the A07.2-era pin by additionally removing the A09 line, and pin the live-master digest in the all-pins test instead of the duplicated fdaf890.

Everything else in the rebase is sound: the descriptor-relevant neighbours from the merged master are handled, the dispatch and help literal are consistent, and the bolt surface is what I approved earlier.

Verdict: approved

The pin-chain fix does exactly what the review asked, and it is minimal:

• pre_a05_descriptor() is now SCHEMA_DESCRIPTOR minus the two A05 edits (the bolt properties line and the connection rewording), which is precisely the descriptor of the released master — the mapping every store created since the native A09 merged carries. It is checked first in pin_accepted, so edit-path capture accepts those stores again.
• pre_a09_descriptor() is redefined as the pre-A05 descriptor minus the A09 line and the Mark change item, which yields the A07.2-era descriptor — so the older pins (the A07.2-era fdaf890, the D1/D2 pin derived from it, and the constant list) are all computed from the real historical chain instead of through the phantom never-shipped mapping.
• The pre-A05 test pins the recomputed live-master digest (085a4f27...), the D1/D2 test derives through the corrected chain with its pinned value unchanged, and the misleading "A05 predates A09" comment is gone.

The recomputed digest is self-validating (the test hashes the constructed descriptor), and by construction the construction equals the released master descriptor given the branch adds only the two A05 edits. The regression surface is closed.