# V3 step 3.4 — records + WITH (done 2026-09-20)

Suite 56/56 (23 run — incl. 3 trap-aborts — + 33 reject).
LL(1)-clean (only the known-benign TypeBlock warning).

## Layout: pointer fields (locked amendment)

The plan said array headers sit inline in records, but 3.1 shipped
nested *separate* objects — one type would then have two layouts
(silent corruption on cross-assigns). Locked instead: records are
flat blobs with scalars/sets inline, array fields as 8-byte
pointers to static descriptors, nested records inline, static
declaration-order offsets. Whole-record assignment is deep
(per-field copy, arrays via `CopyArray`), so no aliasing.

## Grammar (`M2.atg`)

- `RecordType` (+ `RecField`/`RecIdents` reusing the field
  machinery, so classes share offsets for free); class fields
  widened to full `Type` for the same reason.
- `.` suffixes mixing freely with `[...]`
  (`r.tags[i]`, `m[i].x`); 215 non-record, 216 unknown field.
- `KindField` designators resolve through the WITH stacks
  (base + static offset, array fields auto-dereferenced like
  nested levels); field use outside `WITH`/methods is still 230.
- `WITH v [, ...] DO` (215 on non-records): pushes fields
  (inner wins) + bases, pops both exactly; nested and
  field-designator (`WITH r.org`) forms work.
- `Assign` accepts field targets and copies records deep
  (`CopyRecord`); whole-class assign is explicit 230; record/set
  `=` is 213 (added to `EqCheck`); partial composites flow as
  addresses in `Fact` (never scalar-loaded).

## SymTab (additive)

`TypeSize` (inline footprint: arrays/pointers 8), lazy
declaration-order `FieldOffset` (+ rank for declaration-order
enumeration, invalidated on new fields), `FieldOwner` via a new
scope→record link, `FieldCount/FieldName`; nominal-cycle depth
guard (223 proper still deferred).

## Backend (`QbeGen`)

`DeclRec` (statics pre-pass + flat items; reverse-chain emission
walks declaration order via ranks), `FieldAddr`, WITH base stack,
`CopyRecord` (scalars/sets inline, arrays deep, records recurse).

## Bugs caught

1. **Reversed emission**: the field chain builds prepend (reverse
   declaration) but offsets are declaration-order — first cut
   emitted backwards while addressing forwards (SIGSEGV via a
   NULL slot read as pointer). Fixed with declaration ranks.
2. **Inline-vs-pointer size split**: `TypeSize` counted full array
   sizes inside records while layout stores pointers — offsets
   went wrong whenever an array field preceded others. `TypeSize`
   is now the inline footprint (arrays 8) everywhere.
3. Silent partial-composite loads/stores (record/set/class
   through scalar paths) closed by construction; whole-class
   assign plugged with an explicit 230.

## Tests

`t_record` → 160 (nesting, array fields, deep-copy independence
via post-copy mutation, field-designator `WITH`), `t_with` → 114
(multi-designator shadowing, nested `WITH`); `t_bad_field` (216),
`t_bad_with` (215), `t_bad_recassign` (210). Total 56.

Deferred: variant records (Wirth has none — correctly absent),
`WITH`-field `VAR` actuals (needs calls, step 4), class lowering
(vtable/`THIS`), `HIGH`/`LEN`, 223 cycle errors.