spec/ebnf: macroexpand order, set!, letfn primitive, numeric tower

Bring the formal definition in line with this session's language work:
- grammar.ebnf: numbers are a real tower (exact integer / Ratio / double); the M
  suffix reads a real BigDecimal, N an exact integer (drop the stale Janet note).
- 02-reader S5: M is a real java.math.BigDecimal with scale-insensitive equality.
- 03-special-forms: document the read -> macroexpand -> analyze order (macros
  expand before special-form dispatch); special-form heads are not shadowable but
  macros are and value-position locals may be named like a special; set! on a var
  sets the innermost binding (else root); letfn is a primitive with letrec*
  semantics.

docs/grammar.ebnf

@@ -50,11 +50,12 @@ collection      = list | vector | map ;
 nil             = "nil" ;
 boolean         = "true" | "false" ;
 
-(* Numbers. Jolt accepts Clojure's numeric literal syntaxes, but — since Jolt
-   numbers are Janet ints/doubles — it has no distinct bignum, ratio or
-   BigDecimal types: the BigInt suffix N and BigDecimal suffix M are read as the
-   plain number, a ratio a/b is read as its double quotient, and radixed
-   integers are computed by base. The symbolic floats ##Inf/##-Inf/##NaN are
+(* Numbers. Jolt carries a real numeric tower (JVM parity): an integer literal
+   reads as an exact integer (arbitrary precision), a ratio a/b as an exact
+   Ratio, a decimal/exponent literal as a double. The BigDecimal suffix M reads
+   as a real BigDecimal (unscaled x 10^-scale) — 1.5M, 0.0M, 3M; class is
+   java.math.BigDecimal. The BigInt suffix N reads as an exact integer. Radixed
+   integers are computed by base; the symbolic floats ##Inf/##-Inf/##NaN are
    also read. (No octal-with-leading-0 literal.) *)
 number          = symbolic-value
                 | [ sign ] , ( radix-int | ratio | hex-int | decimal ) ;
@@ -62,10 +63,10 @@ sign            = "+" | "-" ;
 integer         = digit , { digit } ;
 hex-int         = "0" , ( "x" | "X" ) , hex-digit , { hex-digit } , [ "N" ] ;
 radix-int       = integer , ( "r" | "R" ) , alnum , { alnum } ; (* base 2..36: 2r1010, 16rFF, 36rZ *)
-ratio           = integer , "/" , integer ;                     (* read as a double quotient *)
+ratio           = integer , "/" , integer ;                     (* exact Ratio *)
 decimal         = integer , [ "." , digit , { digit } ] , [ exponent ] , [ num-suffix ] ;
 exponent        = ( "e" | "E" ) , [ sign ] , digit , { digit } ;
-num-suffix      = "N" | "M" ;          (* BigInt / BigDecimal in Clojure; plain number in Jolt *)
+num-suffix      = "N" | "M" ;          (* N = exact integer (BigInt); M = BigDecimal *)
 symbolic-value  = "##Inf" | "##-Inf" | "##NaN" ;
 digit           = "0".."9" ;
 hex-digit       = digit | "a".."f" | "A".."F" ;

docs/spec/02-reader.md

@@ -43,7 +43,10 @@ exponent := [eE] ['+'|'-'] digits
 - S5. Trailing `N` (BigInt) and `M` (BigDecimal) suffixes are part of the
   grammar; their value semantics are the §4 numeric-tower question.
   Implementations without those towers SHOULD read them as the nearest
-  numeric type and MUST document the choice.
+  numeric type and MUST document the choice. The Chez host carries the full
+  tower: `N` reads as an exact integer (arbitrary precision) and `M` as a real
+  BigDecimal — `1.5M`, `0.0M`, `3M` — with value equality ignoring scale
+  (`1.0M = 1.00M`), `(class 1.5M)` ⇒ `java.math.BigDecimal`, and `decimal?` true.
 
 ### Symbols and keywords
 

docs/spec/03-special-forms.md

@@ -17,7 +17,17 @@ syntax is specified here, their behavior is host-defined.
 
 Special-form head symbols are not shadowable: a binding named `if` does not
 change the meaning of `(if ...)` in operator position. ⚠ This matches the
-reference; it differs from Scheme.
+reference; it differs from Scheme. A local *may* legally be named like a special
+form and used in value position (`(let [if 5] if)` ⇒ `5`); only operator
+position is reserved. **Macros, unlike special forms, ARE shadowable** by a local
+(`(let [when (fn ...)] (when 1 2))` calls the local).
+
+A list form in operator position is resolved in this order (the canonical
+read → **macroexpand** → analyze pipeline): a local binding shadows everything;
+otherwise a macro head is expanded and the result re-analyzed; otherwise a
+special-form head is parsed by rule; otherwise the form is a function
+application. Macroexpansion therefore happens *before* special-form dispatch, so
+a macro is never mistaken for a special form (and vice versa).
 
 ---
 
@@ -135,10 +145,11 @@ S1–S3, E1 → jolt `forms-spec` let group; clojure-test-suite
 | `do` | empty `(do)` → nil; top-level `do` splices for compilation units (important and under-documented) |
 | `fn*` | arities, variadic `&`, closure capture, self-name, simple-symbol params only, recur target |
 | `loop*` | recur arity must match bindings; recur rebinds in place |
+| `letfn` | mutually-recursive local fns (`letrec*` semantics — a fn body sees every binding, not only earlier ones). jolt treats `letfn` as a primitive special, not the reference's `letfn` macro → `letfn*` indirection; behavior is identical |
 | `recur` | tail-position rule (normative definition of tail position needed), across `if`/`do`/`let*`/`try` interactions |
 | `quote` | self-evaluation table: which literals are self-evaluating unquoted |
 | `var` | `#'` reader sugar; resolution at compile time |
 | `throw` | any value vs Throwable — host question; jolt/cljs allow data, reference requires Throwable → classification needed |
 | `try/catch/finally` | catch dispatch order, `:default`-style catch-all is a dialect extension (⚠ divergence note), finally evaluation guarantees, value of try |
-| `set!` | host-dependent (dynamic vars + host fields) |
+| `set!` | `(set! *var* val)` sets the var's innermost thread binding, else its root, and returns val (implemented); a local (deftype mutable field) or host `(.-field obj)` target is host-dependent |
 | `.` / `new` | syntax only; behavior host-defined |