[AntonOresten]

reinterpret(Float4_E2M1FN, 0x13) sets bits above the format's four. Twiddles and table lookups already masked them, but other code read the whole byte:

• show and cvt_generic indexed the tables generated by @microfloat with the whole byte, so they threw BoundsError.
• isnan, iszero, issubnormal, abs, nextfloat, == and others treated those bits as part of the value.

Changes

• Reads: every read of a microfloat's bits goes through an internal raw_bits(x), which masks to bitwidth(T). reinterpret(Unsigned, x) returns the same.
• Constructions: every construction goes through from_raw_bits(T, bits). These two helpers are the only places a later switch to primitive types declared with their real width (Julia 1.14, core: support odd-bit primitive integers, add Core.bitsizeof JuliaLang/julia#61359) would need to change.
• CUDA extension: its private lane_bits was the same mask, so it now uses raw_bits.
• Docs: the Microfloat docstring says the bits above the width are ignored. The exception is ===, which still compares the whole byte.

Behavior change: nextfloat/prevfloat at ±floatmax

On formats without Inf, these stepped past ±floatmax:

• into bits above the width (Float6_*);
• onto the sign bit (Float4_E2M1FN, so nextfloat(floatmax) was -0.0);
• to NaN (the FN formats).

They now stop at ±floatmax, the same way they stop at ±Inf on formats that have it. For the FN formats this changes a valid result (it was NaN, now floatmax), matching Base, where nextfloat never steps from a finite value to NaN.

Testing

• New test/high_bits.jl covers every sub-byte type in the suite:
  • setting bits above the width changes no result;
  • no operation produces such bits;
  • nextfloat/prevfloat stop at the ends.
• Against the code before this PR, the new file gives about 1,600 failures and 2,900 errors.
• Full suite passes locally: 83,470/83,470 on Julia 1.10, 1.13 and 1.14.0-alpha1. That's 50,596 existing tests plus the new ones.
• Generated code is unchanged for valid values. LLVM for the conversions and checks I compared matches Faster microfloat conversions: generated bit-twiddles and branch-free narrowing #35 instruction for instruction, including the vectorized widening loop.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.04%. Comparing base (46108af) to head (feb4cd0).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #36      +/-   ##
==========================================
+ Coverage   86.03%   87.04%   +1.00%     
==========================================
  Files          14       14              
  Lines         659      664       +5     
==========================================
+ Hits          567      578      +11     
+ Misses         92       86       -6     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.