Custodial update, realigning docs and tests

[HuiJun]

Summary

Upstreams the work done on the JPL-Devin/Systemica fork since it was resynced from Open-MBEE/Systemica@97b5edd: 41 commits, 8 already-reviewed-and-merged changes on the fork, fast-forwardable (fork is 41 ahead / 0 behind). Six of the eight are semantic or parser correctness fixes; the corpus gate moves 81/100 → 88/100 clean as a result.

Resolution and semantics (the substance)

• perform / references contribute members. A usage that references another feature (perform a;, ref x ::> y) now contributes that feature's members to resolution, instead of the reference being a dead end.

• satisfy usages are not typed by the referenced requirement. The type checker was reporting false positives on legal satisfy/SatisfyRequirementUsage models; per the spec the usage's type comes from the standard library element, not from the referenced requirement.

• Implicit redefinition of behavior/step parameters is positional, not name-based (KerML 7.4.7.2/7.4.7.3). The old name-matching guard is deleted rather than patched, and a test that encoded the name-based behavior was rewritten:

  - if declaredName == inheritedName { redefine(inherited) }
  + redefine(inheritedParameters[position])   // names need not match
  

  A parameter that a declaration cannot legally redefine now stays inherited, and an implicitly redefining parameter keeps its standard-library base.

• if/else branch bodies are namespaces, so features declared in a branch resolve inside it.

• Message payload features declared in an of clause are registered, so flow of Payload … payload members resolve.

Notation

• Textual notation for history pseudostates (history / shallow history / deep history), deferred events (defer E1, E2;) and entry/exit point pseudostates, wired through the parser, lowering and the runtime. The OMG textual grammar has no production for any pseudostate or for deferral, so these are documented as a Systemica extension in docs/grammar/README.md, not as standard notation. point is matched contextually rather than reserved, since models routinely declare features named point.

Test coverage and docs

• gRPC Layer 2 conformance harness (internal/grpc/conformance_test.go): two fixtures were checked in but read by no test. Now 5 fixture pairs cover Evaluate / Instantiate / ExecuteAction / ExecuteState, asserting the protobuf oneof arm and payload, slots, outputs, state visitation and expected errors.
• docs/SPEC_COMPLIANCE.md and docs/TRAINING_EXAMPLES.md recounted against reality, with a per-file verdict recorded for every corpus expectation change (training_examples_expected.txt was adjudicated file-by-file, never regenerated).

macOS distribution

• docs/MACOS_DISTRIBUTION.md documents the actual cause of the install prompt: it is com.apple.quarantine, not a missing signature — Go's linker already ad-hoc signs darwin/arm64 even when cross-compiled from Linux (verified flags=0x20002 in the Mach-O CodeDirectory of a binary built the way build-release does), so ad-hoc codesign -s - in CI would fix nothing. Removing the prompt entirely requires Developer ID signing + notarization, recorded as a decision record with the exact credentials and runner it needs.
• Release artifacts gain SHA256SUMS.txt and per-platform bundles; packaging/homebrew/ carries a formula template and scripts/render-homebrew-formula.sh. Homebrew is the interim low-friction path (Homebrew-installed binaries are not quarantined) and targets a tap the maintainer will create; nothing here creates or assumes that repo exists.

Verification

CI was unavailable on the fork, so every one of the 8 changes was gated locally, and the merge result at 6bd3ea4 was re-verified end to end:

gofmt -l .          # clean
go build ./... && go vet ./...
go test ./...       # pass
go test -race ./... # pass
make lint           # 0 findings
./scripts/download-training-examples.sh
go test ./internal/core/model/ -run TestTrainingExamples   # 88/100 clean (12 files, 29 errors)

Note that the corpus gate does not run in CI on either repo — neither workflow downloads the corpus, so TestTrainingExamples silently skips there. The 88/100 above is local evidence; adding the download to CI is a recommended follow-up.

Other gates at 6bd3ea4: stdlib parser conformance 94/94, execution conformance 51 cases, golden execution traces 22, runtime robustness 29 subtests, gRPC conformance 5 cases, golden AST fixtures 30, negative parser subtests 32.

Known issues in this branch

• docs/SPEC_COMPLIANCE.md / docs/TRAINING_EXAMPLES.md state 81/100 in places; the later merges took the corpus to 88/100 and the recount has not landed yet.
• The remaining 29 corpus errors span 12 files. 5 of them across 3 files are bugs in the OMG material itself (start/done vs startShot/endShot; the alternative typo), so 95/100 is the material ceiling without changing upstream files.
• The 3 errors in Verification Case Definition Example are likely ours, not the file's: its private import VerificationCases::*; sits inside the verification def body and imports owned by a non-package namespace do not appear to be applied. The checked-in verdict currently blames the file and needs re-adjudication.
• entry <actionName>; (an entry action by reference, rather than an inline body) still fails to parse: expected a body member.