[devin-ai-integration]

What and why

Step 5 of the execution-owned IR migration (docs/internals/execution-ir-design.md, "Migration as a stack of pull requests", row 5), stacked on #2137 (feat/ir-program-ids). Merge order: #2137 first; this PR is then retargeted to develop.

Calc lowering leaves the runtime. runtime.calcShape — the flattened invocation interface the runtime built from symbol.Decl in invoke_calc.go — becomes program.Calc, built by lower.ToCalc. The runtime wraps the record in loadedCalc (the state only the runtime has: the compiled tier and its eligibility, the tool binding, member-declaration metadata, per-run statement orders) and keeps memoizing by symbol in Model.calcShapes / Model.predicateShapes. compileBatch.callees is keyed by *program.Calc; the batch keeps a *program.Calc → *loadedCalc map (held) for the runtime state it settles.

What program.Calc holds and what it defers

Per the design's "Expressions" section, program imports nothing from internal/syntax/ast (the layering test pins it) and expressions are not lowered until steps 6a/6b. So:

• Holds (internal/ir/program/calc.go): the calc symbol, its Origin, qualified Name, Kind and Label; the specialization Chain; the positional Params (name, owner, declaring symbol, the CalcMember declarations along the chain, subject/actor/calc flags); ParamNames and the redefinition Aliases; the Outputs (name, owner, declarations, result/in-out/initial flags); BodyOwner and BodyOutputs; Returns, StepsReturn, StepCount.
• Defers behind program.ExprID: every expression — a parameter's Default, an output's Value, the Result expression — is an opaque id (dense from 1, 0 = none) into a per-calc table.
• Stays on the lower side (lower.CalcTrees): the body statements and their StatementOrder, the steps, the body Nodes and Bindings, and the ExprID → ast.Node table (CalcTrees.Expr). The runtime reads an expression through loadedCalc.Expr(id) exactly where it read the ast.Node field before, so evaluation, error timing and error text are unchanged.

Public surface added

internal/ir/program: Calc, CalcParam, CalcOutput, CalcMember, ResultOutputName, Canonical, AliasRedefined; methods Calc.ResultOutputIndex/ResultOutput/HasInitialOutput/IsCase/Performs/BodyScope/Parameter/HasParameter/SubjectParameter/PositionalIndexes/PositionalArity/OutputNames/UsageSpelling.

internal/ir/lower: CalcEnv, CalcTrees (+ Expr), ToCalc, ToPredicate, ToCalcSignature, CalcChain, CalcKind, CalcStatesComputation, UnboundResultHint, DefaultValueOf, IsCalcDecl, IsPredicateDecl, PredicateKind, IsCalcUsageSymbol, UnwrappedDeclMembers, EachUnwrappedDeclMember, SubjectDecl, SubjectDeclaration.

Nothing from #2137's API changed.

Where the design met the code

• Predicates (invoke_predicate.go) shared calcShape; they lower through lower.ToPredicate, so the runtime keeps one record type and no second AST-reading constructor.
• The library-performance signature (resolveLibraryPerformance) built a bodiless calcShape; that is lower.ToCalcSignature.
• The row names "tests/perf calc benchmarks", but no benchmark in tests/perf matches -bench Calc; the calc-invoking benchmarks are BenchmarkREPLEvalExpr, BenchmarkGRPCEvaluate and BenchmarkConnectEvaluateHTTP (each evaluates Perf::Calc1(2.0, 3.0)), reported below.
• The spec-compliance sources cite file:function pairs (a hygiene test checks them), so the calc rows now cite lower/calc.go, program/calc.go and loadedCalc.

Baselines

None moved: no conformance .expected.json, golden trace, testdata/compiled file, corpus ratchet or pilot-differential record changed.

How it was verified

Full gate on the head (base feat/ir-program-ids merged in with --no-ff): go build ./..., go vet ./..., gofmt -l . (empty), make lint, go test ./..., go test -count=1 ./tests/corpus -run 'TestTrainingExamples|TestPilotCorpora' with both corpora fetched, go test -count=1 ./tests/hygiene, go test -C tools ./..., and the two-line-comment check over the diff (prints nothing). New unit tests in internal/ir/lower/calc_test.go pin the chain/interface lowering, the chained-assignment rule and the dense, deterministic expression ids.

Benchmarks (go test ./tests/perf -run '^$' -bench 'REPLEvalExpr|GRPCEvaluate|ConnectEvaluateHTTP' -benchmem, benchstat, base feat/ir-program-ids vs this head):

Three interleaved comparisons on the same box (n=6, n=8, n=4); benchstat finds no difference the noise does not explain, and the allocation profile is byte- and count-identical:

                      │ feat/ir-program-ids │     this head      │
                      │       sec/op        │  sec/op   vs base  │
REPLEvalExpr-8               9.107m ± 40%     9.621m ± 24%   ~ (p=0.798 n=8)
GRPCEvaluate-8               23.77µ ± 25%     25.34µ ± 44%   ~ (p=0.959 n=8)
ConnectEvaluateHTTP-8        139.1µ           143.9µ         ~ (p=0.886 n=4)
                      │        B/op         │   B/op    vs base  │
REPLEvalExpr-8              6.811Mi ± 0%     6.811Mi ± 0%   ~ (p=0.234 n=8)
GRPCEvaluate-8              10.39Ki ± 1%     10.39Ki ± 1%   ~ (p=0.721 n=8)
ConnectEvaluateHTTP-8       22.85Ki          23.07Ki        ~ (p=0.486 n=4)
                      │      allocs/op      │ allocs/op vs base  │
REPLEvalExpr-8                 110.0 ± 0%      110.0 ± 0%   ~ (all samples equal)
GRPCEvaluate-8                 129.0 ± 0%      129.0 ± 0%   ~ (all samples equal)
ConnectEvaluateHTTP-8          214.0           214.0        ~ (all samples equal)

The wall-clock spread (±25–44%) is the shared runner, not the change: the earlier n=6 comparison put this head at −13% geomean, this one at +6%, both inside the interval.

Pilot execution differential (go run -C tools ./cmd/pilot-exec-diff, evaluator provisioned): agree 229 · kind-only 1 · order-only 0 · disagree 38 · differs-by-design 5 · pilot-unevaluated 167 · pilot-silent 24 · pilot-error 7 · ours-error 10 · ours-undetermined 17 · both-error 16 · nondeterministic 0 (514 cases) — identical to the committed record in docs/project/pilot-execution-referee.md.

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 — internal refactor, no user-visible change
• baselines regenerated and make docs-counts run if a gate count moved (compliance rows need nothing: the census is counted at docs build) — nothing moved
• No internal work-item labels (waves, slices, F4, K5) in the body, docs, or changelog

Link to Devin session: https://nasa-jpl-demo.devinenterprise.com/sessions/d00fc0ddec284f26b9cb865143821490
Open in Devin Desktop: https://nasa-jpl-demo.devinenterprise.com/desktop/session/d00fc0ddec284f26b9cb865143821490?variant=devin
Requested by: @HuiJun

---

🟢 Devin Review found nothing to fix · View findings

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.