Fix Dragon4 shortest formatting for exact powers of two

[tannergooding]

Summary

Number.Dragon4 computed hasUnequalMargins by comparing the extracted mantissa against 1U << TNumber.DenormalMantissaBits. 1U is 32-bit, and C# masks the shift count to 5 bits for a 32-bit operand, so for double (DenormalMantissaBits == 52) this evaluates to 1U << (52 & 31) == 1U << 20, not 1UL << 52.

As a result hasUnequalMargins was wrongly false for every exact power of two in double, so Dragon4 used equal rounding margins and produced a non-round-trippable shortest string for cases like 2^-25 and 2^-958:

Value	Before	After
2^-25	2.980232238769531E-08	2.9802322387695312E-08
2^-958	4.104536801298376E-289	4.1045368012983762E-289

The "before" strings parse back to the adjacent lower-magnitude double.

The fix widens the shift to 1UL, which is correct for all supported types (float 23, double 52, Half 10, BFloat16 7 � all < 64).

---

This is a regression from #102683, which merged the per-type double/float/Half paths into one generic method. The pre-refactor double path used 1UL << DiyFp.DoubleImplicitBitIndex; the merge kept Half's 1U, which silently truncates for double. ExtractFractionAndBiasedExponent and DiyFp.GetBoundaries were unaffected (both already use 1UL).

Testing

• Added DoubleTests.ToString_ExactPowerOfTwo_Roundtrips covering �2^-25 and �2^-958. It fails without the fix and passes with it.
• Existing double/float/Half/BFloat16 ToString tests pass (checked runtime).
• Exhaustive power-of-two sweep (2098 values) and ~10M randomized float/double round-trips, plus exhaustive Half, all pass.

Note

GitHub Copilot helped create this PR.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
13 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-numerics
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

Fixes a correctness issue in the generic Number.Dragon4 floating-point shortest-formatting path where the implicit-bit comparison used a 32-bit shift (1U << ...), producing the wrong result for double and causing non-roundtrippable output for certain exact powers of two.

Changes:

• Widened the implicit-bit check to use a 64-bit shift (1UL << TNumber.DenormalMantissaBits) when computing hasUnequalMargins.
• Added a regression test covering positive/negative double exact power-of-two cases that previously formatted to a non-roundtrippable shortest string.

Show a summary per file

File	Description
src/libraries/System.Private.CoreLib/src/System/Number.Dragon4.cs	Fixes hasUnequalMargins computation by using a 64-bit shift to correctly detect the implicit bit for double.
src/libraries/System.Runtime/tests/System.Runtime.Tests/System/DoubleTests.cs	Adds regression coverage ensuring affected exact power-of-two double values format and parse round-trip as expected.

Copilot's findings

• Files reviewed: 2/2 changed files
• Comments generated: 1

[Copilot]

Copilot's findings

• Files reviewed: 2/2 changed files
• Comments generated: 0 new