Take the UID registry out of callers' hands, and stop building it twice.

[xbmlz]

The defect

uid exported three package-level maps: Dictionary (496 registered UIDs with their PS3.6 metadata), KeywordToUID (494 keyword → UID), and Known (one Info per registered UID, filled by a package init). Read-only by convention, mutable by type, no synchronisation of any kind.

uid.Known[uid.ImplicitVRLittleEndian] = someOtherInfo

That is one assignment, from anywhere in any dependency, and every implicit-VR file in the process decodes as explicit from then on. Same defect #81 fixed for the data dictionary — this time in the table that decides how bytes are read.

The change

Approved API shape: uid.Lookup + uid.LookupKeyword.

// UID -> what the registry records, plus the encoding facts that follow from it
info, ok := uid.Lookup(uid.CTImageStorage)

// keyword -> UID (the direction the old Lookup went)
u, ok := uid.LookupKeyword("CTImageStorage")

• Lookup(u UID) (Info, bool) is what replaces the maps for readers. Info is a strict superset of a dictionary entry plus the four derived transfer-syntax flags, and it is exactly what the three internal call sites already wanted — DecodeDatasetContext, determineWriteEncoding and EncodeDataset all read IsTransferSyntax / IsImplicitVR / IsLittleEndian off it. It returns a copy, so the table is not reachable through the result.
• LookupKeyword(keyword string) (UID, bool) is new, and is the old Lookup's job under a name that says which direction it goes.
• Dictionary, KeywordToUID and DictEntry are unexported — in generate_uid_dict.py as well as in the generated file, so a regeneration cannot quietly export them again.

Two deviations from a literal "unexport everything"

1. Known is deleted, not unexported. Every field of an Info is either read straight from the dictionary entry or derived from the UID in three comparisons, so Lookup derives them on demand. That is one table where there were two, nothing to keep in step, and no init work — 496 Info values built at program start — for a program that never looks a UID up.
2. The root aliases godicom.UIDDictionary and godicom.KnownUIDs are deleted, not unexported. An alias to another package's identifier cannot be unexported, and there is nothing left for them to name. godicom.UIDInfo and godicom.LookupUID are unchanged.

Both are stated in the CHANGELOG.

Evidence

Three mutations, each restored and re-verified:

mutation	fails
info.IsImplicitVR = false	TestLookup + 8 root-package tests
unregistered UID answered true	TestLookup, TestDetermineEncodingPrivateTransferSyntaxRequiresArgs
Keyword: ""	TestLookup, TestLookupCoversEveryEntry, TestLookupKeyword

The first found a real gap — nothing had pinned the implicit-VR true case, only the false ones — which TestLookup now covers.

TestLookupCoversEveryEntry walks the whole table and checks each entry resolves with every field intact, which is the invariant len(Known) == len(Dictionary) used to state. TestLookupReturnsCopy mutates the returned Info and checks the registry and UID.Name() are untouched, so a future memoising *Info cache cannot regress the point of the change.

Local gate green: build, go vet, gofmt, staticcheck -checks=all, go test ./... on amd64 and under GOARCH=386 CGO_ENABLED=0, all eight cross-vet targets. Lookup, LookupKeyword and entry at 100%; ExtraInfo/IsRetired lifted 66.7% → 100%.

Downstream

Breaking, which 0.x allows. Within the organisation the only affected code is two lines in one gonetdicom test — ae/storage_sop_classes_test.go:35 and :121 — recorded in the CHANGELOG for whoever bumps the module there. No non-test code anywhere in the org touched the maps.

🤖 Generated with Claude Code