[devin-ai-integration]

What and why

AssignmentLoopStep allocates 971 KB per step on the 0.12.0 candidate against 921 KB on 0.11.0's source built with the same toolchain (six rounds, bytes exact; allocation counts are equal). Half of that is the size of EvalContext, which the runtime copies for every nested evaluation: #1514 added chainStatic with two slice headers (the chain's segments and the declarations they name, 48 bytes), and #1461 added the openInputs flag in its own padded word, so a context grew from 152 to 184 bytes.

This PR keeps what both merges compute and shrinks how the context notes it:

• chainStatic holds a pointer to the chain's last segment, its length and a pointer to the memoized chainDecls instead of two slice headers; chainDeclarationsFor compares the last segment's address as before, so the whole-chain / suffix distinction is unchanged. The Model.chainDeclarations memo stores *chainDecls, so the memoized slice is shared, not copied, into each context.
• openInputs sits beside the behavior-body flag so the two share a word.

EvalContext is 152 bytes again. Root cause: the per-context copy of chain state (#1514) and padding (#1461); correctness of both merges is kept.

A/B, six rounds (go test ./internal/exec/runtime -bench . -benchmem -count 1 × 6, benchstat; pre-#1514 / #1514 / candidate 1c6bb1972 / this branch; same host, "Intel(R) Xeon(R) Processor", 8 CPUs; bytes and allocations exact)

Row	pre-#1514	#1514	candidate	fix	0.11.0 source, go1.27.3
AssignmentLoopStep B/op	912.7 Ki	974.1 Ki (+6.73 %)	947.9 Ki (+3.86 %)	925.7 Ki (+1.42 %)	899.7 Ki
AssignmentLoopStep allocs/op	8 536	9 335	8 539	8 539	8 531
DerivedReadWriteRead B/op	2.656 Ki	2.750 Ki (+3.5 %)	3.250 Ki (+22.4 %)	3.188 Ki (+20.0 %)	2.6 Ki
DerivedReadWriteRead allocs/op	67	67	71	71	67
every other row of the package, B/op and allocs/op				within ±0.2 % of the candidate

Residual, stated in the fragment: AssignmentLoopStep stays 13 Ki above pre-#1514 (the context's new state — the chain note and the open-inputs flag — is what the merges needed); the remaining DerivedReadWriteRead bytes are #1619's order marks on derived reads, a separate PR (perf/derive-marks-footprint). No sec/op claim: the host ran a full go test ./... gate alongside rounds 1–3 and the timing columns carry ±100–250 % spread.

Memory acceptance

runtime.test -test.bench '^BenchmarkAssignmentLoopStep$' -test.benchtime 300x, candidate and fix built the same way, three cold runs each, interleaved (/usr/bin/time, 8 CPUs):

run	candidate peak RSS	wall	fix peak RSS	wall
1	117.5 MiB	0.82 s	112.7 MiB	0.81 s
2	115.4 MiB	0.82 s	119.6 MiB	0.82 s
3	119.8 MiB	0.83 s	115.9 MiB	0.82 s

Hard cap: under ulimit -v 4000000 both pass (candidate 122 MiB, fix 118 MiB peak RSS). Under ulimit -v 2000000 both revisions die as the Go runtime reserves its arenas (runtime: out of memory: cannot allocate 4194304-byte block (24 MB in use)), before any model is loaded, so that cap says nothing about the change. Whole binary, sysml -validate over the 12 086-element generated model, three cold runs: 198 / 198 / 198 MiB candidate, 200 / 187 / 200 MiB fix, 0.24 s both; tests/stressmodel FleetInstantiate/512 and Load/512 stay inside their GC-pacing spread (1.14–1.47 GiB and 1.27–1.36 GiB on both). Resident memory is not what this change recovers: it is 23 KB of allocation volume per step, and the PR does not claim an RSS change.

How it was verified

• TestChainDeclarationsAreResolvedOncePerModel (new, internal/exec/runtime/chain_declarations_memo_test.go) pins that a chain's declarations are resolved once per model and that the contexts evaluating it share the memoized entry.
• The execution conformance and trace suites, robustness suites and the gRPC conformance cases pass unchanged (go test ./... below).

Gate

$ go build ./...        # ok
$ go vet ./...          # ok
$ gofmt -l .            # (empty)
$ go test -timeout 45m ./...
ok   124 packages, 0 FAIL, exit 0

(corpora present: training examples, pilot corpora, spec examples, pilot library XMI, PSSM suite, with their OPENSYSML_REQUIRE_* variables set.)

Checklist

• go build ./..., go vet ./..., gofmt -l . (empty), go test ./... pass locally (above); make lint not run (the gate above is the repository's definition of done)
• Tests added or updated for the change
• Documentation extended where it already covers the surface — none covers EvalContext's layout
• Changelog entry added as changes/unreleased/eval-context-size.fixed.md
• baselines regenerated and make docs-counts run if a gate count moved — no count moved
• No internal work-item labels in the body, docs, or changelog

Link to Devin session: https://nasa-jpl-demo.devinenterprise.com/sessions/15408827f8104c7fbbdbd1b2c852ed2c
Open in Devin Desktop: https://nasa-jpl-demo.devinenterprise.com/desktop/session/15408827f8104c7fbbdbd1b2c852ed2c?variant=devin
Requested by: @HuiJun

[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.