docs(adr): add API stability contract for float8 v1.0.0

Author: dndungu

docs/adr/001-api-stability-v1.md

@@ -0,0 +1,93 @@
+# ADR-001: API Stability Contract for float8 v1.0.0
+
+- **Status:** Accepted
+- **Date:** 2026-03-29
+- **Authors:** Daniel Ndungu
+
+## Context
+
+The `github.com/zerfoo/float8` package provides IEEE 754 FP8 E4M3FN arithmetic for the Zerfoo ML ecosystem. It is imported by `ztensor` for quantized tensor storage and compute. The package has reached a stable API surface and needs a clear stability contract so downstream consumers can depend on it without fear of breakage.
+
+## Decision
+
+### Stable (v1 guarantee)
+
+The following API surface is covered by Go module compatibility and will not have breaking changes within the v1.x line:
+
+**Core type:**
+- `Float8` (defined as `uint8`)
+
+**Constructors and conversions:**
+- `ToFloat8(float32) Float8`
+- `ToFloat8WithMode(float32, ConversionMode) (Float8, error)`
+- `FromFloat64(float64) Float8`
+- `FromInt(int) Float8`
+- `FromBits(uint8) Float8`
+- `Parse(string) (Float8, error)`
+- `Zero() Float8`, `One() Float8`
+
+**Methods on Float8:**
+- `ToFloat32() float32`, `ToFloat64() float64`, `ToInt() int`
+- `Bits() uint8`
+- `Abs() Float8`, `Neg() Float8`
+- `Sign() int`
+- `IsZero() bool`, `IsNaN() bool`, `IsInf() bool`, `IsFinite() bool`, `IsNormal() bool`, `IsValid() bool`
+- `String() string`, `GoString() string`
+
+**Arithmetic functions:**
+- `Add`, `Sub`, `Mul`, `Div` (and `*WithMode` variants)
+- `AddSlice`, `MulSlice`, `ScaleSlice`, `SumSlice`
+
+**Math functions:**
+- `Sqrt`, `Pow`, `Exp`, `Log`
+- `Sin`, `Cos`, `Tan`
+- `Floor`, `Ceil`, `Round`, `Trunc`, `Fmod`
+- `Min`, `Max`, `Clamp`, `Lerp`, `Sign`, `CopySign`
+
+**Comparison functions:**
+- `Equal`, `Less`, `LessEqual`, `Greater`, `GreaterEqual`
+
+**Batch conversions:**
+- `ToSlice8([]float32) []Float8`
+- `ToSlice32([]Float8) []float32`
+
+**Configuration:**
+- `Config`, `DefaultConfig`, `Configure`
+- `ConversionMode` (constants: `ModeDefault`, `ModeStrict`, `ModeFast`)
+- `ArithmeticMode` (constants: `ArithmeticAuto`, `ArithmeticAlgorithmic`, `ArithmeticLookup`)
+- `EnableFastConversion`, `DisableFastConversion`
+- `EnableFastArithmetic`, `DisableFastArithmetic`
+- `DefaultConversionMode`, `DefaultArithmeticMode` (package-level variables)
+
+**Constants:**
+- Bit masks: `SignMask`, `ExponentMask`, `MantissaMask`, `MantissaLen`
+- Exponent: `ExponentBias`, `ExponentMax`, `ExponentMin`, `Float32Bias`
+- Special values: `PositiveZero`, `NegativeZero`, `PositiveInfinity`, `NegativeInfinity`, `NaN`, `MaxValue`, `MinValue`, `SmallestPositive`
+- Math constants: `E`, `Pi`, `Phi`, `Sqrt2`, `SqrtE`, `SqrtPi`, `Ln2`, `Log2E`, `Ln10`, `Log10E`
+
+**Error types:**
+- `Float8Error` (struct with `Op`, `Value`, `Msg` fields)
+- Sentinel errors: `ErrOverflow`, `ErrUnderflow`, `ErrNaN`
+
+**Utilities:**
+- `Initialize()`, `GetVersion()`, `GetMemoryUsage()`, `DebugInfo()`
+
+### Explicitly deferred
+
+The following are **not** part of v1 and are candidates for v1.1+:
+
+- **FP8 E5M2 format** — A second 8-bit format with 5 exponent bits and 2 mantissa bits, used in some gradient representations. Will be added as a separate type (e.g., `Float8E5M2`) without altering the existing `Float8` (E4M3FN) type.
+- **SIMD-accelerated batch operations** — Platform-specific vectorized paths for slice operations.
+- **Stochastic rounding mode** — A `ConversionMode` variant that uses probabilistic rounding for training workloads.
+
+### Versioning policy
+
+- Patch releases (v1.0.x): bug fixes, performance improvements, documentation.
+- Minor releases (v1.x.0): new functions, types, or constants that do not break existing callers.
+- The `Version` constant tracks the current release and is updated by release-please.
+
+## Consequences
+
+- Downstream packages (`ztensor`, `zerfoo`) can pin `float8 v1.x` and upgrade freely within the major version.
+- New FP8 formats (E5M2) will be additive and will not modify the `Float8` type or its semantics.
+- Any behavioral change to existing functions (e.g., rounding rules, special-value handling) requires a new major version.