# V3 self-hosting fixpoint reached

Tag `v3-fixpoint`. Main suite 111/111.

## The milestone

V3, **compiled by V3 itself**, compiles the whole compiler source set to
a **byte-identical** `gen_ssa/M2.ssa`:

```
M2   (built by gm2 -fiso)  compiles  SymTab QbeGen M2S M2P M2(driver)
                          ->  gen_ssa/M2.ssa         (1,991,439 bytes)
M2self (built by V3 from that image)  compiles the same set
                          ->  gen_ssa/M2.ssa
cmp: *** IDENTICAL (FULL FIXPOINT) ***
```

So Stage 2 (V3 compiling itself) reproduces Stage 1's image exactly —
the self-hosting fixpoint.

## The last fix

`QbeGen.InitHeap`, for a nested array (`ARRAY … OF ARRAY …`) on the
heap/stack, allocated each element sub-array with `NewHeap` but never
**initialised its descriptor header**. A local `ARRAY [0..15] OF Name`
therefore had sub-descriptors with garbage headers, and passing
`names[i]` to an `ARRAY OF CHAR` formal trapped in the scanner's
`GetName`. Fix: call `InitHeap(eb, elem)` on each sub-array.

Repro:

```modula2
TYPE Name = ARRAY [0..63] OF CHAR;
PROCEDURE P;
  VAR names: ARRAY [0..3] OF Name;  i: CARDINAL;
  PROCEDURE Set (VAR s: ARRAY OF CHAR); BEGIN s[0] := 'x' END Set;
BEGIN i := 0; Set(names[i]) END P;
```

## How we got here (this arc)

From "sources compile and link" to the fixpoint, each step a real V3
runtime/ABI bug with a minimal repro: nested modules; inline record
array fields; the NUL-terminated string model; `p^` as a `VAR` actual;
de-eagered NIL guards; off-by-one string scans; short-circuit `AND`/`OR`;
and now nested-array heap initialisation.

## Remaining (bootstrap polish)

- `bootstrap/fixpoint.sh`: automate Stage1 gm2 → Stage2 V3 → Stage3 V3
  and `cmp` the images (the manual chain above already shows Stage2 ==
  Stage1). Stage-aware CR flag: `-m` for gm2, no `-m` for V3.
- Cross-stage `FileIO` selection (stage 1 uses the host `FileIO`; stage 2+
  the self-hosted `runtime/syslib/FileIO`).

## Files

`compiler/src/{QbeGen.mod,M2.lst}`.