[#175 architectural follow-up] Tokenizer-level fix for Variable :-naivete (compound-label-in-USE-args)

[msarson]

Surfaced at #175 Phase A. Deferred to architectural review.

Context

#175 shipped option (b) diagnostic-side guard in AttributeDiagnostics.validateAttributeApplicability to suppress the bare-identifier Variable:Attribute false-positive surface (e.g. USE(RCFilter_SL_Clients:External)). The diagnostic-side fix is narrow + targeted; works correctly. But it leaves the 3-token split intact in the token stream (`Variable(MyVar) + Delimiter(:) + Attribute(External)`), which is grammatically incorrect for compound USE-label names per Clarion convention.

Architectural question

Is the diagnostic-side workaround the right long-term shape, or do we want to invest the 30-file cascade audit + type-keyword-list synchronization burden to land the tokenizer-side fix?

Phase A analysis (from #175)

• 128 references to TokenType.Variable across 30 files (vs FieldEquateLabel's 4 files — [Diagnostics] Attribute-applicability false positive on :Suffix label names (any attribute keyword) #174's cascade was much smaller). Architectural-surprise threshold hit per Mark's pause-rule.
• Substrate-asymmetry framing does NOT transfer cleanly from [Diagnostics] Attribute-applicability false positive on :Suffix label names (any attribute keyword) #174. [Diagnostics] Attribute-applicability false positive on :Suffix label names (any attribute keyword) #174's Variable-was-the-outlier framing worked because Label already includes `:`. Here, Variable's `:`-naivete is grammatically intentional because `:` is the type-annotation separator (sibling `TypeAnnotation` pattern). Making Variable `:`-greedy with negative-lookahead on type keywords is regex-complexity addition, NOT substrate-asymmetry restoration.
• Type-annotation regression risk: load-bearing. `MyVar:BYTE` must still tokenize correctly. Negative-lookahead in Variable regex needs to share the type-keyword list with the existing TypeAnnotation regex — duplication risk + drift risk.

Three candidate fix layers

(a) Tokenizer-level Variable extension with negative-lookahead

Sketch from #175 issue body:
```ts
[TokenType.Variable]: /\b(?!(?:IF|LOOP|...))[A-Za-z_][A-Za-z0-9_](?::(?!\s(?:byte|short|ushort|long|ulong|real|sreal|decimal|pdecimal|string|cstring|pstring|date|time)\b)[A-Za-z_][A-Za-z0-9_]*)?\b/i
```

Root cause fix. Requires:

• Full 30-file cascade audit (any consumer that walks neighboring tokens around Variable could break)
• Type-keyword list synchronization between Variable regex + TypeAnnotation regex (likely factor into shared constant)
• Comprehensive test coverage on type-annotation edge cases

(b) Already shipped at #175

Diagnostic-side guard in AttributeDiagnostics. Narrow scope; 0 other Attribute consumers affected per audit. Adequate for the immediate user-visible surface.

(c) New CompoundLabel TokenType

Cleanest separation but adds new TokenType. Tier-walker + switch-case-by-type cascade across all consumers. Over-engineered for current surface.

Recommendation

Defer indefinitely unless one of these triggers:

• A new compound-label-consuming feature surfaces that requires the single-token shape (e.g. a refactoring tool that needs to rename `PrefixGroup:Field` atomically)
• The diagnostic-side guard pattern repeats in other diagnostics (would suggest the workaround is becoming structural debt)
• Mark calls option (a) the right architectural direction during a deeper review

Current option (b) is fully adequate per the #175 consumer audit; no other Attribute consumer has the false-positive surface.

Cross-references

• MT kanban: `(filled at task-file time)`
• Parent: [Diagnostics #174 follow-up] Attribute-applicability false positive on bare-identifier compound USE-labels (Variable pattern + colon) #175 (commit body has full Phase A analysis)
• Related: [Diagnostics] Attribute-applicability false positive on :Suffix label names (any attribute keyword) #174 (the `?`-prefixed half — different cascade shape, different fix-layer answer)
• Architectural standing rule: Mark's "pause + queue for me when cascade audit reveals tokenizer contract has many fragile consumers"

Priority: low. Not a user-visible blocker.