[devin-ai-integration]

What and why

Since #1441 a submission's parse trees ([]parsed, one parser and tree per file) were passed on past submissionResult to resolveReferences/pullReferences, so they stayed referenced through the analysis of every load, for the root namespaces an unresolved reference would be looked up by. Before #1441 they died once acceptParsed had read them. The 0.12.0 gate measured the cost as the +31 MiB peak resident set of sysml -validate on the 12 086-element generated model against 0.11.0, and bisected it to #1441 by the heap held at exit (56 → 74.5 MB).

resolveReferences and pullReferences now take the submission's []SourceFile and parse again (parseRoots) only when the analysis left a subject unresolved — the rare case pays a second parse; the common case holds no trees through the analysis. workspaceLoad.parses becomes sources. The lazy walk of #1624 is kept: a submission that resolves parses nothing twice and walks nothing (TestResolvingNothingDoesNotWalkTheText, adapted to pass sources).

Fragment: changes/unreleased/release-preparse-trees.fixed.md.

How it was verified

Memory acceptance, bin/sysml -validate built as make build builds it (CGO_ENABLED=0 -ldflags "-s -w", go1.27.3 for every column), one warm-up run per binary and model, then three cold runs per configuration, interleaved; medians, peak RSS in MiB (the three runs in brackets). Columns: 0.11.0 source · the merge before #1441 (511061d46) · #1441 (c082e554c) · develop (1c6bb1972) · this branch.

Workload	0.11.0	pre-#1441	#1441	develop	this branch
generated 12 086, GOGC=100	162 (162,162,162)	163 (157,163,163)	176 (177,176,176)	181 (181,180,194)	177 (177,172,179)
generated 12 086, GOGC=25	146 (146,144,146)	147 (149,147,145)	165 (164,165,165)	166 (165,166,166)	150 (150,150,150)
generated 6 198, GOGC=25	115	114	125	127	118
Apollo 11 (28 files), GOGC=25	152	149	161	163	156
wall, generated 12 086	0.26 s	0.26 s	0.27 s	0.24 s	0.24 s
GOMEMLIMIT=120MiB, generated 12 086	0.28 s · 141	0.27 s · 141	0.46 s · 154	0.44 s · 155	0.25 s · 144
ulimit -v 1200000, generated 12 086	0.26 s · 161, rc 0	0.27 s · 176, rc 0	0.27 s · 173, rc 0	0.25 s · 193, rc 0	0.24 s · 176, rc 0

GOGC=25 keeps the resident set near the live heap and is the row that separates the configurations; at GOGC=100 the collector's pacing spreads single runs by ±10 MiB (develop's 194). Under the memory limit the candidate spends 0.44 s to 0.11.0's 0.28 s and this branch 0.25 s. Under ulimit -v 600000 every binary, 0.11.0 included, fails at start from the runtime's address-space reservation, so 1.2 GB is the cap used.

Six rounds, benchstat, same four revisions, the load rows (bytes and allocations unchanged by this branch, as expected — the trees were garbage either way; what moves is when):

                          │   pre1441    │               m1441                │                dev                 │                fix7                │
                          │     B/op     │     B/op      vs base              │     B/op      vs base              │     B/op      vs base              │
LoadModel/elements=0-8      227.9Ki ± 0%   227.9Ki ± 0%  -0.03% (p=0.002 n=6)   246.8Ki ± 0%  +8.29% (p=0.002 n=6)   246.8Ki ± 0%  +8.29% (p=0.002 n=6)
LoadModel/elements=250-8    13.96Mi ± 0%   13.96Mi ± 0%       ~ (p=0.132 n=6)   14.71Mi ± 0%  +5.38% (p=0.002 n=6)   14.71Mi ± 0%  +5.38% (p=0.002 n=6)
LoadModel/elements=1000-8   55.54Mi ± 0%   55.55Mi ± 0%       ~ (p=0.589 n=6)   58.32Mi ± 0%  +5.01% (p=0.002 n=6)   58.33Mi ± 0%  +5.02% (p=0.002 n=6)
LoadModel/elements=4000-8   222.4Mi ± 0%   222.4Mi ± 0%       ~ (p=0.937 n=6)   233.5Mi ± 0%  +4.99% (p=0.002 n=6)   233.5Mi ± 0%  +5.00% (p=0.002 n=6)
                          │  allocs/op  │  allocs/op   vs base                │  allocs/op   vs base              │  allocs/op   vs base              │
LoadModel/elements=0-8      2.022k ± 0%   2.022k ± 0%       ~ (p=1.000 n=6) ¹   2.155k ± 0%  +6.58% (p=0.002 n=6)   2.155k ± 0%  +6.58% (p=0.002 n=6)
LoadModel/elements=1000-8   533.1k ± 0%   533.1k ± 0%       ~ (p=0.747 n=6)     538.3k ± 0%  +0.98% (p=0.002 n=6)   538.3k ± 0%  +0.98% (p=0.002 n=6)
LoadModel/elements=4000-8   2.113M ± 0%   2.113M ± 0%       ~ (p=0.589 n=6)     2.133M ± 0%  +0.95% (p=0.002 n=6)   2.133M ± 0%  +0.95% (p=0.002 n=6)
                          │    sec/op    │    sec/op     vs base              │    sec/op     vs base              │   sec/op     vs base               │
LoadModel/elements=1000-8   111.6m ± 10%   115.0m ± 19%       ~ (p=0.589 n=6)   106.2m ±  7%       ~ (p=0.093 n=6)   106.3m ± 1%   -4.75% (p=0.002 n=6)
LoadModel/elements=4000-8   482.3m ±  3%   474.0m ± 23%       ~ (p=0.699 n=6)   455.4m ± 14%       ~ (p=0.065 n=6)   458.0m ± 4%   -5.04% (p=0.015 n=6)
REPLLoadFile-8        1.522 ± 4%   1.528 ±  7%       ~ (p=0.485 n=6)   1.408 ± 56%        ~ (p=0.065 n=6)   1.401 ± 5%   -7.94% (p=0.004 n=6)
REPLSubmitSnippet-8   1.180 ± 6%   1.167 ± 26%       ~ (p=0.818 n=6)   1.018 ±  4%  -13.75% (p=0.002 n=6)   1.046 ± 6%  -11.34% (p=0.002 n=6)

(The +5 % B/op of develop over pre-#1441 on every load row is #1478, addressed in its own PR; the LoadModel/0 allocations are #1462, likewise.)

Definition of done on this branch:

$ go build ./... && go vet ./... && gofmt -l .      # clean, gofmt prints nothing
$ go test ./...                                      # every package ok, with two caveats resolved below:
FAIL  tests/identity   TestPilotLibraryXMI: OPENSYSML_REQUIRE_PILOT_LIBRARY_XMI=1 but pilot library XMI not downloaded
FAIL  tests/corpus     TestPSSMSuiteMigration: OPENSYSML_REQUIRE_PSSM_SUITE=1 but the PSSM suite is not downloaded
FAIL  cmd/sysml        panic: test timed out after 10m0s (TestExploreTablesTheSpacecraftRaceWithinItsBudget at 5m19s) — run while six benchmark rounds loaded the machine
$ ./scripts/download-pssm-suite.sh && ./scripts/download-pilot-library-xmi.sh
$ go test -count=1 ./tests/identity -run TestPilotLibraryXMI && go test -count=1 ./tests/corpus -run TestPSSMSuiteMigration
ok    github.com/Open-MBEE/OpenSysML/tests/identity   0.638s
ok    github.com/Open-MBEE/OpenSysML/tests/corpus     2.666s
$ go test -count=1 -timeout 40m ./cmd/sysml
ok    github.com/Open-MBEE/OpenSysML/cmd/sysml   233.784s

The two corpus gates failed only because the measurement VM set the require variables without holding those corpora; fetched, both pass. Nothing in this branch touches identity or migration.

Checklist

• make test and make lint pass locally — go build ./..., go vet ./..., gofmt -l . empty, go test ./... above
• Tests added or updated for the change — TestResolvingNothingDoesNotWalkTheText now drives the sources path
• Documentation extended where it already covers the surface (see CONTRIBUTING.md) — none covers the retention
• 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 — no gate count 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/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