Widen pseudo for numerical work: sized types, FLOAT32, exact built-ins, USE IEEE

Pseudo could say what a procedure computes only up to INTEGER and FLOAT,
so it could not state what a Mojo or Python program with Int32, UInt8 or
Float32 values will compute, and it had no words for rounding itself.

- lang/parse: INT8..INT64, UINT8..UINT64, FLOAT32, FLOAT64 as types
  (INTEGER or FLOAT with a width, so every "whole number" rule accepts
  them); conversions written as Mojo writes them, INT32(X), FLOAT32(X);
  1_000_000, 0xFF and 0b1010 literals; a type name as an argument
  (MAX_VALUE(INT8)); the USE line (USE IEEE now, library modules next).
- check: widths and precisions never mix silently (mixed-width,
  mixed-precision); a literal takes the type of the sized value beside
  it, as in Mojo, and one that does not fit is an error before any run
  (literal-range); typing for every new built-in.
- interp: a sized result, store, parameter, RETURN or conversion that
  does not fit stops the run (kind overflow), WRAP(X, T) wraps on
  purpose; FLOAT32 results are rounded to single precision; USE IEEE
  lets NaN and infinities flow.
- numeric.js: exact helpers: FMA rounded once, EXACT(X) with every digit
  a FLOAT holds, ULP and NEXT_AFTER for both precisions, bit operations
  in two's complement.
- Built-ins: ASIN ACOS ATAN SINH COSH TANH HYPOT POW CBRT EXPM1 LOG1P E
  TRUNC SIGN COPYSIGN CLAMP FMA IS_NAN IS_INF IS_FINITE INF NAN ULP
  NEXT_AFTER IS_CLOSE EXACT MAX_VALUE MIN_VALUE EPSILON GCD LCM BIT_AND
  BIT_OR BIT_XOR BIT_NOT SHIFT_LEFT SHIFT_RIGHT WRAP. They stay usable
  as names (a variable E, a FUNCTION GCD): the program's own wins.
- vcgen: an overflow obligation on every sized result, then the range
  as a fact (also for sized parameters and loop-cut variables); FLOAT32
  arithmetic as its own uninterpreted functions, constants rounded
  exactly, TOFLOAT32(v) otherwise; argument obligations for SQRT, the
  logarithms, ASIN/ACOS, LOG1P, CLAMP and the shifts; nothing proved
  under USE IEEE.
- testing: the fuzzer tries the edges of a sized type and generates
  FLOAT32 values.
- pseudo-to-mojo: Int8..Float32, conversions, the new built-ins through
  std.math (each name in the harvested index), a note where Mojo wraps.
- Tests: numeric.test.mjs; corpus 49-52; prover fixtures int8-overflow,
  float32-store, ieee-nan; the reference documents every new form.

Gate: 324 pass (core). New Mojo uses std.math names the interpreter has
not been verified on (listed in sync-mojo's KNOWN_GAPS); they need
pixi run mojo-verify on the network before the interpreter runs them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FgvbRdoEheAHqsRmr2WuKZ
This commit is contained in:
alexmandClaude committed 2026-10-05 11:43:53 +00:00
1 parent c5157fcad5
commit be0db2d74f
24 files changed
+1271 -96

No files matched your search

+22
View File
@@ -20,6 +20,25 @@ The prover is three TypeScript modules under strict mode; the rest of the engine
"Proved" is about the notation's own number model. INTEGER is a whole number of any size and a FLOAT is a finite number whose rounding the prover does not see into. A run of the interpreter stops where a value leaves that model (a whole number past 2^53 − 1, a FLOAT that is not finite, a division by zero), so every run stays inside what a proof covers. The Mojo made from a program has other limits (a 64-bit `Int` that wraps, a `Float64` that becomes `inf` or `nan`), and a unit whose proof leans on the number model carries an `int-model` or `float-model` note saying so.
### Numbers
Pseudo is meant for numerical work, so its number model is stated exactly ([the reference](../src/common/reference.ts) shows every form with a runnable example):
| Type | Holds | When a result does not fit | The prover |
|---|---|---|---|
| `INTEGER` | A whole number of any size; a run holds it exactly to 2^53 − 1 | The run stops past 2^53 − 1 | Exact arithmetic of any size; an `int-model` note says the lab's and Mojo's limits are not part of the proof |
| `INT8` … `INT64`, `UINT8` … `UINT64` | Two's complement or unsigned, n bits (64-bit types held exactly to 2^53 − 1 in this lab) | The run stops (kind `overflow`); `WRAP(X, T)` wraps on purpose (up to 32 bits) | Exact arithmetic plus an `overflow` obligation on every result, store, parameter, RETURN and conversion: "fits INT8". After it, the range is a fact. A sized parameter and a sized variable cut by a loop keep their range |
| `FLOAT`, `FLOAT64` | IEEE 754 double | The run stops on infinity or NaN | `FLOAT+` and its kind are uninterpreted: the same operation on the same values gives the same result, nothing more |
| `FLOAT32` | IEEE 754 single: every result rounded to 24 bits (`Math.fround`, exact for +, −, ×, ÷ of singles) | As FLOAT | Its own uninterpreted functions (`F32+`, ...); a constant is rounded exactly; a value that becomes a FLOAT32 is `TOFLOAT32(v)`, about which nothing is assumed |
Rules the checker holds: two different sized types never mix silently (`mixed-width`, `mixed-precision`; convert with `INT32(X)`, `FLOAT32(X)`); a numeric literal takes the type of the sized value beside it, as in Mojo, and a whole-number literal that does not fit is an error before any run (`literal-range`); INTEGER may go into a sized type (checked when it happens), never a different sized type.
`USE IEEE` lets NaN and the infinities flow as IEEE 754 says (SQRT(−1.0) is NaN, 1.0 / 0.0 is infinity; whole-number division by zero still stops). The prover's reasoning about numbers does not hold for those values, so in such a program nothing is marked proved, and each unit carries an `ieee` note.
The numerical built-ins (`HYPOT`, `EXPM1`, `LOG1P`, `FMA`, `ULP`, `NEXT_AFTER`, `EXACT`, `IS_CLOSE`, `MAX_VALUE`, the bit operations, ...) are exact or correctly rounded: `src/engines/pseudo/numeric.js` computes `FMA` with one rounding by exact binary arithmetic, `EXACT(X)` gives every decimal digit a FLOAT holds, and `ULP` and `NEXT_AFTER` step through the representation. Functions JavaScript has no exact version of (erf, gamma) are not built-ins; they belong in the library as verified pseudo. Built-ins that stop the run for some arguments (SQRT, the logarithms, ASIN and ACOS, LOG1P, CLAMP, the shifts) carry an `argument` obligation. The built-ins added for numerical work stay usable as names (a variable `E`, a `FUNCTION GCD`): the program's own definition wins (`SOFT_BUILTINS` in `lang.js`).
`tests/numeric.test.mjs` pins each rule; the corpus programs `49-sized-integers` to `52-elementary` show them in use.
### Soundness
A "proved" must be a proof. Two test files hold the prover to that, and a change to `logic.ts`, `solver.ts` or `vcgen.ts` must keep both green:
@@ -36,6 +55,9 @@ When a soundness bug is found, add the program that shows it to `tests/prover/`
| `fill-negative.pseudo`, `randint-empty.pseudo` | `FILL` and `RANDINT` assumed facts about arguments the interpreter checks by stopping, without asking whether they hold | An `argument` obligation, and guarded facts |
| `float-overflow.pseudo` | The prover treats FLOAT as numbers (`Y == Y`), while a run went on with infinity and NaN | The interpreter stops on a FLOAT that is not finite |
| `foreach-empty.pseudo` | The way out of `FOR EACH` kept "X is in the collection", false after zero passes over an empty list | The fact holds only inside a pass |
| `int8-overflow.pseudo` | (a guard written with the sized types) Sized arithmetic treated as exact with no range would prove `X + 1` fits INT8 | Every sized result has an `overflow` obligation |
| `float32-store.pseudo` | (a guard) A whole number stored in a FLOAT32 keeping its exact value would prove `X == N` for N = 16777217 | A value becoming a FLOAT32 is rounded (constants) or `TOFLOAT32(v)` |
| `ieee-nan.pseudo` | (a guard) `X == X` for every number fails for NaN | Nothing is proved in a program that says `USE IEEE` |
In the app, an obligation the prover marks proved that a generated input breaks at the same place is shown as broken, not proved.