# Step: VAL conversions (LONGINT/REAL) + long→int narrowing

Tag `v3-val-conversions`. Suite **148/148**; fixpoint **OK**
(image **2,320,770 bytes**).

## What

`VAL` now covers the numeric conversions V3 lacked:

| From → To | Codegen |
| --- | --- |
| `INTEGER` → `LONGINT` | `extsw` (sign-extend) |
| `LONGINT` → `INTEGER` | `copy w` (low 32 bits) |
| `INTEGER` → `REAL` | `swtof` |
| `LONGINT` → `REAL` | `sltof` |
| `REAL` → `INTEGER` | `dtosi` (truncate toward zero) |
| `REAL` → `LONGINT` | `dtosi` |

(Ordinal↔ordinal, pointer↔pointer and REAL↔REAL were already allowed.)

New `QbeGen` helpers: `NarrowLong`, `ConvRI`, `ConvRL`, `ConvLR`
(alongside the existing `WidenLong`/`ConvIR`).

## Payoff

`Conversions` is simplified: `StrToInt`/`StrToCard` now accumulate in
a `LONGINT` and simply range-check, then narrow with `VAL(INTEGER, v)`
/ `VAL(CARDINAL, v)` — the previous "detect overflow before the 32-bit
wrap" hand-guard is gone.

## Tests

`t_valconv.mod` (15): int→long→int round-trip, int→real, real→int
truncation, long→real. `ConvProg`/`ConvResultsProg` still pass with
the rewritten `Conversions`.

## Files

`compiler/src/QbeGen.def`/`.mod` (`NarrowLong`/`ConvRI`/`ConvRL`/
`ConvLR`), `compiler/src/M2.atg` (`VAL` action),
`stdlib/conversions.mod`, `compiler/tests/t_valconv.mod`,
`compiler/run_tests.sh`, `docs/features.md`.