perf(resolve): match names first under recursive imports

[devin-ai-integration]

What and why

Since the hiding of colliding imported memberships (KerML 7.2.5.4) landed, a lookup of one name through import X::** enumerated every element below X and ran the per-element hiding checks (admitsUnderName, reexportedUnder, hiddenOnRoute, withoutHiddenMembers) on each of them before comparing names. On models built from recursive imports (the OMG Vehicle Example in examples/pilot-corpora/sysml-examples) name resolution became about sixty percent slower, and the tests/model race shard sits near its timeout.

The enumeration in internal/semantic/resolve/filter.go now takes a memberSelect:

• named: a lookup lists, per visited namespace, only the members bound to the name — the scope's LookupLocalAll plus derived short names (shortNamedMembers), the child imports asked for the same name, and the index entries bound to it (declared name, short name, or a derived short name for an entry the index alone holds). The hiding, re-export and visibility checks run on those alone.
• namespaces: the recursive descent lists only the members that can be namespaces (own members or imports, or index entries under their name; never a body-local scope) and checks only those, so a namespace that is hidden or not admitted still hides its subtree as before.
• the zero value lists every member as today; ImportedElements, completion, the Duplicate of imported member name audit and the distinguishability warnings are unchanged.

Around it, none changing an answer:

• derivedShortName memoizes a settled answer per symbol (journaled to the symbol's document frame), skipped while a naming chain under way cut it short, while the lookup was cut by an import cycle, and while a filter condition resolves unfiltered or an import all lifts the visibility boundary.
• A scope's derived-short-name chains are resolved for a name only when the scope has derived-named members at all. Gating them further on Index.ShortNamed was tried and dropped: its segment read made a recorded document's analysis stale as soon as another document re-exported its names, so the record was hydrated (TestWorkspacePortLintReadsRecordedDefinitions).
• Index.GetFQN caches the built FQN per symbol for the index generation (FQNOf was ~27% of all allocation: the string was rebuilt for every namespace a lookup visited).
• elementList dedups by a linear scan until it holds sixteen keys, then indexes.

memberCache.fqn joins the lazily filled caches TestSnapshotIndexMatchesFreshLoad skips when comparing a decoded snapshot with a fresh load.

Specification basis

No behaviour change. KerML 7.2.5.4 / 8.3.2.4.3 hiding of indistinguishable imported memberships (declared and derived short names, KerML 8.2.3.3) is applied exactly as before, to the candidates instead of to the whole subtree; Resolver.InCondition is untouched. tests/resolve gains TestDerivedShortNameFromADeclarationUnderAnUnnamedOwner: a redefinition borrowing a short name declared under an anonymous part, which no index entry holds.

How it was verified

Measured on one machine (8 cores), go test -count=1 -run 'TestInterfaceRecordDifferentialCorpora/sysml-examples' ./tests/model, back to back (wall):

	non-race	-race
develop before the hiding change (0af9fcb36)	69.3 s	737.5 s
develop at 6b81872 (before #1155, #1158 and the day's merges)	99.5–108.8 s	670.7 s
this branch on that develop	82.2 s	635.3 s
develop at e00ce72 (with #1158)	110.6–117.3 s	—
this branch merged with that develop	114.2–116.3 s	—

On current develop the costs this change removes no longer show in the profile (subtree enumeration per lookup, derived-short-name chains: under 1 s), but the subtest's time is now dominated by Model.generalizationTarget (41 s of 160 s CPU, through collectContributors and generalsOf) and by the captured collision sets of #1158 (importedCollisions → capturing → collectImported, 28 s, mostly import-target resolution and metadataBodyOwner ending in generalization resolution), so the end-to-end number does not move there. The -race column was taken in one sequence on the earlier develop.

Passing: internal/semantic/..., internal/check/..., internal/workspace/..., internal/frontend/..., tests/resolve, tests/export, tests/migrate, tests/parser, make lint, and tests/model TestIncrementalEqualsFresh, TestSnapshotIndexMatchesFreshLoad, TestInterfaceRecordDifferential$, TestPackagesImportingEachOtherInACycleAnalysePromptly.

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

Link to Devin session: https://nasa-jpl-demo.devinenterprise.com/sessions/26443c7e5db24d5c99c23cd94846f6a0
Open in Devin Desktop: https://nasa-jpl-demo.devinenterprise.com/desktop/session/26443c7e5db24d5c99c23cd94846f6a0?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]

Closing: on current develop the subtest this change targeted (TestInterfaceRecordDifferentialCorpora/sysml-examples) measures the same with and without it (develop 110.6–117.3 s, branch merged with develop 114.2–116.3 s); the costs it removed no longer show in the profile, which is now dominated by Model.generalizationTarget and the captured collision sets. The export-shard promptness failure seen here is fixed separately in #1198.