# V3 step 1.10 — Clarion classes: decl + inheritance + impl blocks (done 2026-09-19)

Pre-step-3 grammar work from `docs/OOP.txt` (757-line Clarion CLASS
spec + Table AVL example). Suite 33/33 (14 run + 19 reject).
Grammar stays effectively LL(1)-clean (one benign Coco warning,
see below).

## Locked decisions (interactive §5)

- Classic identifiers stay: no underscores, so Table's `_Find` etc.
  remain lexically out of reach (noted, not solved).
- Method separator `;` — the Table example overrules the sketch's `,`.
- Construct now, lowering later (units/procs/nesting pattern).
  Full `Table.mod` needs steps 3+4+stdlib (pointers, calls, SIZE…).
- `CLASS IMPLEMENTATION` + single `(Parent)` inheritance included
  now (parsed, scoped, 230); multiple parents → 230.

## Grammar (`compiler/src/M2.atg`)

- One `ClassItem = "CLASS" ("IMPLEMENTATION" … | …)` in both the
  `TYPE` loop and `DeclSeq`: splitting decl/impl across the two
  loops needs 2-token lookahead (`CLASS ident` vs
  `CLASS IMPLEMENTATION`), which LL(1) cannot do.
- `ClassRest`: name + optional parents + `;`, `{field ";"}`,
  `{method ";" + CloseProc}`, `END`-match, one 230. Fields reuse
  the record machinery (`FieldPending/FixPendingF`); `x = Name`
  constant form included.
- `Parents`: first parent resolved (201 unknown, 230 non-class,
  stored via `SetParent`); extras consumed + 230 (single
  inheritance locked).
- `MethodHeading = [VIRTUAL] ProcHeading` (+ `MarkVirtual`);
  bodies reuse `Block` (nested procedures inside methods work,
  as in Table's `Search`/`_Find`); impl-side existence check
  (`MethodExists`, else 201); `FORWARD` allowed in both.
- `ClassImplRest`: class resolution (201/230), `PushClassMembers`,
  methods, optional `BEGIN` body, `END`-match, one 230.
- Field *use* (`KindField` in Design) is now explicit 230 instead
  of silent `0` — no silent acceptance anywhere anymore.

## SymTab (additive)

`FClass`/`ClClass` (=11); `NewClass`, `SetParent`, `PushClassScope`
(records the body scope on the descriptor), `MethodExists`,
`PushClassMembers` (fields only — see bug 1), `MarkVirtual`
(`virt` node flag); `FieldPending`/`PushRecord` accept `FClass`;
methods stay `KindProc` (class membership rides `node^.scope`).

## Bugs caught

1. **Member-copy collision**: first cut re-entered sibling methods
   into impl member scopes, so every method body heading hit
   `200 duplicate` (+ `232` cascade via cleared `curProc`).
   Fix: members push fields only; sibling calls honestly 201
   until step 4.
2. **Coco's lingering TypeBlock warning** (`CLASS` start+successor):
   verified benign — the generated `WHILE`/`IF` decision is correct
   by inspection and by behavior on every input shape. Coco is
   strict about `FIRST ∩ FOLLOW` on that loop; the code it emits
   does the right thing. Documented, not "fixed".
3. **A phantom `t_class` failure** (`ident expected` at the impl
   line) appeared during development across many consistent runs,
   then vanished and never reproduced (deterministic 3/3 + full
   suite since, including from a clean rebuild). Token-stream dumps
   of failing vs working inputs were byte-identical through the
   failure point. No mechanism found; no code changed to "fix" it.
   If it ever recurs: clean-rebuild first, then suspect toolchain
   (CR codegen) rather than the grammar — the decision code reads
   correctly. Recorded here so nobody re-bisects it blind.

## Tests

`t_classdef` (fields, `VIRTUAL`, empty/full formals, result →
2× 230: class + unit), `t_class` (inheritance + impl + body with
`RETURN 0` → 4× 230, zero cascades), `t_bad_parent` (unknown
parent → 201 + 230). Total 33 (14 run + 19 reject).

Deferred: vtable/dispatch/`THIS`, multi-inheritance, opaque
classes, `SHORTINT [-1..1]` postfix bounds, full Table (steps 3+4).