# Step 11 — Separate compilation units (DEFINITION/IMPLEMENTATION/IMPORT)

157/157 tests green (144 + 7 run + 6 rejection), mc64 boot + example
green, Showcases unchanged (157, 83, 168), new Showcase4 prints 225
as hand-computed.

## Goal

`DEFINITION`/`IMPLEMENTATION` compilation units with `IMPORT`/`FROM`,
compiled from several files into a single `.MC4` image (the remaining
big item from the step-10 limits).

## What was built

- Driver (`compiler.frm`, the true source of generated `M2c.mod`):
  multi-file session `M2c lib.def libimpl.mod ... prog.mod`, one
  `.LST` per input, fail-fast on the first bad file (no HALT abort:
  a clean `EXIT` with `Incorrect source`). `SymTab.Init` and
  `MGen.OpenModule("")` run once; the program unit sets the image
  name (`MGen.SetModName`).
- Grammar (`M2c.atg`): `M2c = Unit`, `Unit = DefUnit | ImplUnit |
  ProgUnit`. New `DEFINITION`/`IMPLEMENTATION` keywords were added in
  end-of-file productions so existing token numbers only shifted by
  position — CR regenerates `M2c.mod`'s `Msg` table from `M2c.err`
  (verified identical mapping); new-productions note added.
- `SymTab`: definition interfaces publish to the existing export
  tables plus an in-memory side table (8 defs, 32 names each);
  `OpenImplementation` re-enters the interface into a fresh scope
  (slots allocated once at definition, found via the module
  context); procedure headings match through the forward machinery
  (`ReuseProc`/`VerifyProc`, 231 on mismatch); `CloseImplementation`
  requires every body (231). `FROM L IMPORT x` materializes real
  alias symbols (`EnterImpProc` shares the proc number, so call
  checking is unchanged; VAR/CONST aliases resolve through the new
  `GlobAlias` hook in `MGen.QualGlob`, the single funnel for all
  global use-paths). Ghosts are 201, bad kinds 221, duplicates 200.
- `run_tests.sh`: `expect_run_files` / `expect_fail_files` for
  multi-file sessions.

## Tests — 157/157 (144 + 7 + 6)

7 run: `d_basic` 25 (FROM var/func, qualified const/var/call,
`VAR` param, lib init), `d_from` 12 (exported array type, private
impl helper, `VAR` composite formal), `d_multi` 82 (two libs,
implementation-level `IMPORT`), `d_init` 510 (init order across
files, init reading another lib via `FROM`), `d_func` 55
(recursive lib function), `d_type` 1122 (exported record, `VAR`
record formal), `d_case` 25 (qualified enum `CASE` labels).
6 rejections, each exactly one error: `d_bad_nodef` 201,
`d_bad_sig` 231, `d_bad_nobody` 231, `d_bad_imp` 201,
`d_bad_exp` 201, `d_bad_dup` 200.

## Bugs found and fixed

1. `DefProcHead` double-semicolon: heading already consumes its own
   `;` (`';' expected` on the first definition).
2. `HALT` on compile failure aborts (SIGABRT, `Abandon` noise in
   every rejection run) — replaced with a clean loop `EXIT`.
3. Value composite formals in a test (`Sum(x : Vec)`) correctly hit
   the pre-existing 230 rule — test switched to `VAR` formal.
4. `M2c.mod` is REGENERATED by CR from `compiler.frm` on every
   build: driver edits first went to `M2c.mod` and were silently
   lost. Driver work belongs in `compiler.frm` (same for the `Msg`
   table, auto-mirrored from `M2c.err`).

## Known limits (deferred)

`M.x` as a `VAR` actual is still 233 (pre-existing local-module
limitation, confirmed identical on step-10 build — Showcase4 works
around it via a temp); second program unit rejected (230 + follow-on
200); multi-`.MC4` images with `depCount` loading (still single
image, `depCount` 0); circular imports; opaque types; `WITH`-field
`VAR` actuals; `CHAR`/`BOOLEAN` tails as `VAR` actuals.