feat(values): Go infrastructure for numeric loss signals

[aalpar]

Summary

Lays the Go-side foundation for the no-information-loss numeric pipeline
(design: plans/2026-05-14-numeric-loss-signals-design.md). This is PR 1 of 3.

• New values/ public API: ToFloat64WithAccuracy, ToFloat64Lossless,
  ToComplex128WithAccuracy, ToComplex128Lossless, Complex128Result struct
• New werr.ErrLossyConversion sentinel for precision and imaginary-drop errors
• Accuracy symbols: SymbolAccuracy{Below,Exact,Above} + BigAccuracyToSymbol
• Per-kind helpers for all 7 numeric types registered in NumericTypeSpec
• BigFloat API hygiene: Float64() → Float64Truncated() rename +
  new Float64WithAccuracy() method; 13 call-site migrations in extensions/, parser/, registry/

Key correctness fixes (caught by crosscheck)

Two direction-recovery bugs identified by post-implementation review:

1. Rational overflow-to-Inf (values/rational.go): when a huge rational
   overflowed to ±Inf, (*big.Rat).SetFloat64(±Inf) returned nil and the
   function fell through to return f, big.Exact, true. Fixed: detect math.IsInf(f,0)
   before the round-trip and infer direction from sign.

2. Integer MaxInt64 saturation (values/integer.go): float64(MaxInt64) = 2^63
   (rounds up), then int64(2^63) saturates back to MaxInt64 in Go, making
   the round-trip falsely compare equal. Fixed: early-exit with big.Above when
   f >= 9223372036854775808.0.

Test plan

• 60 subtests in values/conversion_test.go covering all 7 kinds ×
  {exact, below, above, overflow, NaN} × {lossless-ok, lossless-error}
• MaxInt64 and rational-overflow cases explicitly tested post-fix
• make lint && make covercheck both pass (38 packages ≥ 80%)
• go test ./... all green
• Master CI green at time of push (run 25892419247)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR lays the Go-side foundation for loss-signal-aware numeric conversions by threading big.Accuracy (and complex “isReal” signaling) through the values numeric registry and adding public conversion helpers (ToFloat64*, ToComplex128*) plus supporting sentinels/symbols. It also performs a BigFloat API hygiene rename (Float64() → Float64Truncated()) and updates call sites/tests accordingly.

Changes:

• Add ErrLossyConversion, accuracy singleton symbols, and new values public conversion helpers (accuracy + lossless wrappers).
• Replace numeric-registry conversion hooks with To{Float64,Complex128}WithAccuracy and migrate internal call sites.
• Rename BigFloat’s lossy Float64() to Float64Truncated() and introduce Float64WithAccuracy(), updating downstream call sites.

Reviewed changes

Copilot reviewed 30 out of 30 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
werr/werr.go	Adds ErrLossyConversion sentinel.
VERSION	Bumps version string.
values/symbols_accuracy.go	Introduces accuracy singleton symbols + mapper.
values/rational.go	Implements rational float/complex conversion w/ accuracy recovery.
values/promotion.go	Migrates Number→float/complex helpers to new registry API.
values/numeric_tower.go	Updates registry lookup usage for simplify/exactness.
values/numeric_registry.go	Replaces registry conversion hooks; renames LookupNumericSpec → Lookup.
values/numeric_registry_test.go	Updates registry tests for new conversion APIs.
values/numeric_methods_coverage_test.go	Migrates BigFloat call sites to Float64Truncated().
values/integer.go	Adds integer conversion w/ accuracy (direction recovery).
values/float.go	Adds float conversion w/ always-exact accuracy.
values/conversion.go	Adds new public conversion helpers + Complex128Result.
values/conversion_test.go	Adds table-driven tests for new conversion helpers.
values/complex.go	Adds complex conversion w/ ok real-flag + accuracy.
values/big_number_test.go	Migrates BigFloat test call sites to Float64Truncated().
values/big_integer.go	Adds BigInteger conversion w/ big.Accuracy.
values/big_float.go	Renames Float64() → Float64Truncated(); adds Float64WithAccuracy().
values/big_complex.go	Adds BigComplex conversion w/ per-component accuracy; migrates call sites.
values/big_complex_test.go	Migrates BigFloat test call sites to Float64Truncated().
TODO.md	Adds follow-up tracking section for loss-signals API evolution.
registry/helpers/value_conv.go	Migrates BigComplex conversion to Float64Truncated().
plans/2026-05-14-numeric-loss-signals-impl.md	Updates implementation plan to reflect current state/decisions.
plans/2026-05-14-numeric-loss-signals-design.md	Updates design doc (hybrid return shape, BigFloat hygiene, etc.).
internal/parser/parser_test.go	Migrates BigComplex float extraction to Float64Truncated().
internal/parser/parser_number.go	Migrates BigComplex→inexact conversion to Float64Truncated().
internal/parser/parser_number_test.go	Migrates BigComplex parse assertions to Float64Truncated().
internal/parser/big_number_test.go	Migrates BigFloat test call sites to Float64Truncated().
extensions/math/prim_transcendental.go	Migrates BigComplex transcendentals to Float64Truncated().
extensions/math/prim_rounding.go	Migrates BigFloat rounding primitive to Float64Truncated().
extensions/math/prim_complex.go	Migrates BigComplex magnitude/angle to Float64Truncated().

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[aalpar]

Addressed Copilot review in 9d96a56:

1. values/integer.go:175 — comment correction
The original comment said "rounds toward zero" which is true for the specific 2^53+1 → 2^53 case but misgeneralizes. Go's int64→float64 conversion is round-to-nearest, ties to even per IEEE 754. Rewrote the comment with that rule plus a 2^53+3 → 2^53+4 counter-example to forestall the same simplification mistake.

2. values/conversion_test.go — negative-zero detection
Tightened the assertion path in TestToFloat64WithAccuracy: when tc.wantValue == 0, the test now calls math.Signbit on both got and want in addition to qt.Equals. The two test cases that exercise this — float-negative-zero and big-float-underflow-negative — now actually fail if the implementation returns +0 where -0 is expected.

3. VERSION bump
False positive — the pre-commit hook auto-bumps VERSION on every commit in this repo (see memory/feedback-copilot-version-false-positive.md). Reverting would just cause the next commit to bump it again. The release-cut is a separate signed commit; this PR is not it.

The same commit also folds in post-PR-1 cleanup: isAlwaysReal registry shortcut, the NumberToComplex128 → NumberToComplex128Lossy rename for naming-convention consistency, the *WithAccuracy / *Truncated / *Lossy / *Lossless suffix taxonomy in CODING_STYLE.md, and 8 deferred follow-ups in TODO.md — including 5 pre-existing BigComplex precision-loss bugs the Float64Truncated rename made visible.

Bench-gate against master at the cleanup HEAD: geomean -0.62% (HEAD slightly faster) at N=3, well within the per-run noise band (~±1.5%); 10/16 benchmarks favor HEAD. Hot-path arithmetic methods (Add/Subtract/Multiply/Divide on *Integer etc.) do not reference the new helpers — verified by grep.