Remove gratuitous const_cast<GenContext&> casts; legalize stop() (A6 phases 0–1)

[ASDAlexander77]

Summary

First implementation slice of the A6 plan (docs/MLIRGen-refactoring-review.md, §3): 11 of the 31 const_cast<GenContext &> casts removed with zero semantic change.

• Phase 1 — stop() legalized. stopProcess is now mutable and stop() is const: it's an out-of-band cancellation signal, which is exactly what mutable is for. The internal const_cast<GenContext *>(rootContext) and both casting call sites disappear.
• Phase 0 (discovered during the census) — nine casts were never mutations:
  • Six sites cast only to call non-const methods on the funcOp op handle (getSymName, getCallableResults, getCallableRegion, setPersonalityAttr, debug print). MLIR op handles are cheap value types — copying the handle gives a non-const handle to the same operation, so these now copy instead of casting. (setPersonalityAttr mutates IR, not the context; behavior identical.)
  • *(const_cast<GenContext &>(genContext).inferTypes) — the member is a pointer; reading it from a const context yields pointer-to-non-const already.
  • getConditionalType cast a local, non-const GenContext.
• Also removed a stale commented cast referencing a nonexistent funcOpVarScope field, and updated A6 in the review doc with the phase 0/1 status and remaining census — everything left (thisType, typeAliasMap, loop flags, generatedStatements, typeParamsWithArgs, MLIRCodeLogic.h) is a genuine mutation for phases 2–4.

Test plan

• tslang rebuilt cleanly (header change rebuilds all dependent TUs)
• All seven JIT smoke tests produce identical output
• MLIRGenTests.exe: 12/12 pass

🤖 Generated with Claude Code