feat(migrate): run configurations, property-valued probabilities, draw policies and a result comparison harness

[devin-ai-integration]

What and why

The SysML v1 migrator wrote a decision edge's «Probability» as a constant even when the tag names a property of the analysis block, so every migrated run drew a fixed tier mix instead of the one a Cameo run configuration sets; the SimulationProfile:SimulationConfig elements themselves (run count, duration mode, time variable, execution target, result location) and the tool's stored result snapshots were reported unmapped, and nothing compared a migrated configuration's -runs distribution with those snapshots.

• Property-valued «Probability». activity.go: probability() resolves a tag that names a property (by name or XMI id) visible from the activity or its context block to a feature reference, @Probability { p = ProbabilityBTOOP; }, so the object the behavior runs on decides the weights. Tags naming nothing visible, a private property, a non-property or a non-numeric one are report entries with the reason. Lowering type-checks a feature-valued p against Probability::p, and the [0, 1] / sum-to-one checks made of constants at lowering are made of the read values when the decision is reached, each an ErrBranchWeights; guarded-out branches are filtered before the draw and the rest renormalised.
• Draw policy. runtime.DrawPolicy (random|min|max|average) states how every RandomFunctions call resolves — min/max/average need no seed, normal has no finite min/max and returns a typed error naming the call. Threaded like the seed: -draws, %draws, %runs <n> [<seed>] <action> (seed optional under a fixed policy, still refused under random), analysis wire/JSON ("draws"), external and standing engines, gRPC, witnesses (draws by max) and replay: compatibility. Weighted decisions stay seeded whatever the policy.
• SimulationConfig → v2. simconfig.go writes each configuration, recognised by its profile's provenance, as an action def holding the executionTarget as part target and perform action run ::> target.<behavior> (inherited classifier behaviors resolved), annotated with the new non-normative OpenSysML Libraries/Simulation.sysml metadata @Simulation::Configuration { runs; draws; timeVariable; startTime; stepSize; timeUnit; parallelForks }; every other setting is kept in a comment. Targets the migration did not write, non-part targets, state machines and classifiers with no behavior are reported with the reason. ActionExecutor.Results() also exposes <part>.<attribute> of the parts the action holds so the target's attributes are observable after the run.
• Results and the harness. results.go indexes the tool's result-snapshot instances (classifier-less or typed) under each configuration's resultLocation/result package into a JSON sidecar (-migration-results <file>), resolving observable names through defining features and inferring classifiers from slot owners. Snapshots in a shared result package whose configured feature values differ from what the target configures are excluded with a note rather than mixed in. sysml <migrated>.sysml -compare-results <sidecar> [-action ...] [-runs n] [-draws p] [-seed s] [-observe stored[=feature]...] runs each configuration with its recorded count and policy (or the ones given) and prints tool vs OpenSysML min/mean/p50/p90/max per observable with the relative difference; empty, non-numeric and unrun observables are reported, never dropped.
• Report hygiene. SimulationConfig elements, their target and result instances and the represented «Probability» edges are mapped where written.
• A profile namespace ending in .xmi (SimulationProfile.xmi) is pinned as a profile, not the OMG XMI namespace, by tests in internal/translate/xmi and its sysmlv1 reader.
• The sidecar schema is the new package internal/translate/simresults (standard library imports only), registered in the translate layer of tests/hygiene/layering_test.go: it describes what the migration writes, and migrate, the CLI and the REPL comparison read it through that layer; no permitted or tolerated entry changed.
• Two runtime fixes the configurations needed: a performed feature chain joins the performance its object already runs, and a parallel region stood for by a stateless state needs no initial.
• The self-model, README and testing notes count 104 bundled library files / ten extensions; the pilot-differential baseline differs from develop's in the examples/ digest only, re-recorded for that figure.

TMT proof (TMT-2024x.mdzip)

	total	mapped	approximated	unmapped	skipped
develop	44604	27943	10285	6376	1713
this branch	44604	27967	10261	6376	1713

The +24 mapped / −24 approximated is 24 SimulationConfig classes (of 62, all approximated as plain classes before) and 6 decision nodes becoming mapped, against 6 unmarked control flows beside marked ones now reported approximated with the complementary weight they are given; no other report entry moves. The 38 remaining configurations are approximated with the reason (12 name no execution target, the classifier behavior of the target is written as a verification def rather than an action def, treatAllClassifiersAsActive = false, …). The migrated file validates with 0 errors. 30 of the 62 configurations have stored snapshots (82 total).

Acquire Target with IRIS and NFIRAOS - Logical, seed 1, Time_Acq_Total:

configuration	tool runs	tool mean	OpenSysML runs	OpenSysML min / mean / p50 / p90 / max
Group 0 (recorded 1000 runs)	4	16.46 s	1000	3.00 / 21.33 / 19.22 / 39.76 / 91.83 s
Group 0	4	16.46 s	100	3.16 / 21.55 / 19.90 / 39.17 / 78.75 s
Group 1	4	157.21 s	100	42.50 / 103.05 / 92.40 / 153.27 / 227.57 s
Group 3	1	208.4 s	100	51.12 / 117.70 / 105.58 / 186.81 / 331.51 s
Group 6	1	210.4 s	100	45.97 / 117.71 / 107.63 / 195.81 / 310.44 s

For comparison, develop's migration of the same activity (fixed p = 0.5) gives mean 133.31 s (30.86–315.42 s) at 100 runs. The Group 0 target class as saved in the model configures only ProbabilityAcqPwfsSecond = 1.0; the four stored snapshots recording that configuration are the ones compared, and the nine others in the shared Results::Group 0 package (six recording AcqTelPoint = BTOOP = ChkLaserProp = 1.0, mean 96.6 s, and three recording AcqTTFOIWFS = 1.0) are excluded with a note as runs of another configuration. Under -draws max, Align BTO OP Logical on its target is 90.2 s for every run (min 9.2 s, average 49.7 s).

Known limitations

• The remaining gap to the tool's numbers on the acquisition workflow is the opaque JavaScript retry loop in Acquire and Lock PWFS Logical, which is migrated by another change; it also makes the Group 0 configuration exceed the action-step budget under a fixed draw policy, so the deterministic-duration proof uses Align BTO OP Logical.
• The «SimulationConfig» settings other than the seven metadata fields (executionListeners, animationSpeed, UI options, …) are kept in a comment and are not executable.
• The sidecar's snapshot values are the tool's; a snapshot slot defined outside the document (MonteCarloAnalysis::Mean, …) or holding a non-number is a note, not a value.
• A target property whose default literal is empty configures nothing: it is read as an unspecified value the run fills, so it neither filters snapshots nor is compared.

Specification basis

SysML v2 v2.0 §7.16 (Actions: decision nodes and succession guards) for the weighted-successor semantics; the «Probability» tag, SimulationConfig, result snapshots and durationSimulationMode are MagicDraw/Cameo profile encodings with no OMG counterpart, migrated by provenance. docs/project/spec-compliance.md gains rows for the draw policy and non-constant branch weights; docs/reference/sysml-v1-migration.md gains the mapping rows and a "Run configurations" section.

How it was verified

• gofmt -l . (empty), go vet ./..., make lint, go build ./..., go test ./... (incl. ./tests/...) all pass locally.
• Corpus gates with the corpora downloaded (OPENSYSML_REQUIRE_TRAINING_CORPUS=1 OPENSYSML_REQUIRE_PILOT_CORPORA=1 OPENSYSML_REQUIRE_PILOT_LIBRARY_XMI=1 go test -count=1 ./tests/corpus ./tests/identity -run 'TestTrainingExamples|TestPilotCorpora|TestPilotLibraryXMI|TestCorpusGates|TestCorpusRoundTrip'): 100/100 training clean, pilot ratchets unchanged, 9823 identities agree.
• make docs-check, python3 scripts/changelog.py check, python3 scripts/check-doc-ids.py pass.
• New coverage: migrate fixtures under tests/migrate (testdata/xmi/simconfig.xmi) (property-valued probability, config → action def, min/max/random/average modes, result snapshots incl. classifier-less, nested packages, other-configuration exclusion, empty defaults); conformance stochastic_draws_{min,max,average}, stochastic_draws_max_normal_unbounded, stochastic_weights_read_from_performer, action_part_attributes_in_outcome, action_perform_chain_joins_object_performance, state_parallel_stateless_region (+ trace goldens); robustness_draw_policy_test.go; REPL/CLI/analysis/gRPC/witness tests for -draws/%draws/replay; cmd/sysml/compare_test.go for the harness; internal/translate/xmi namespace tests; go test ./tests/... incl. the layering gate.
• By hand: the TMT migration and runs above; the migrated model validates with 0 errors.

Checklist

• make test and make lint pass locally
• Tests added or updated for the change
• Documentation extended where it already covers the surface (see CONTRIBUTING.md)
• Changelog entry added as changes/unreleased/<slug>.<section>.md, not as an edit to CHANGELOG.md
• baselines regenerated and make docs-counts run if a gate count moved (compliance rows need nothing: the census is counted at docs build)
• No internal work-item labels (waves, slices, F4, K5) in the body, docs, or changelog

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring