# P2 Pascal validator pilot

`p2_pascal.atg` describes Wirth/Ammann Pascal-P2 (1972-73) as
implemented by `pcom`. It was authored from the P4 pilot skeleton and
corrected against some 60 oracle probes — every construct below was
checked against the real compiler, whose parser (`pcomp.pas`) is the
ground truth for the language. (The `Grammars/P2/p2_pascal.atg` file in
the reference collection is a mislabeled Oberon grammar, unrelated.)

## Oracle notes

`pcom` was built with `fpc -Miso` from the ISO-7185-modified sources
(`Pascal-P2-master/source/`), needing a one-character fix in a
`/tmp` copy (duplicate `'1'` in an identifier-scan set that fpc
rejects). Behaviors the harness must respect, all verified:

- **Uppercase only** (CDC charset): lowercase never assigns the
  scanner symbol and the compiler ping-pongs forever. Test sources
  are uppercase; the validator is case-sensitive and rejects
  lowercase cleanly (oracle-consistent, never a false positive).
- **Trailing newline required**: files without one spin at EOF
  (`EOF ENCOUNTERED` flood). The harness appends `\n`.
- **Spin on sync-loss**: illegal characters (`_ & { @ #` → error
  399) and some shapes never recover — the compiler loops instead of
  listing. Harness rule: timeout (10s; valid programs finish
  instantly) counts as oracle-reject. Every accept-case was verified
  to terminate.
- Verdicts: `**** ^NNN` markers in the listing (`NNN` is Wirth's
  error number); absence = accept. Semantic markers (undeclared
  identifiers etc.) also count as reject, so the battery uses only
  well-formed programs.
- `PROGRAM T; BEGIN END.` (no `VAR`) and `PROGRAM HELLO(OUTPUT)`
  both terminate cleanly — used as harness sanity checks.

Key probed P2 facts (vs P4): `AND`/`OR`/`NOT` words but **no `&`**;
`DIV`/`MOD`/`IN`; `CASE` with **no `ELSE`**; `(* *)` comments that
do **not nest**, no `{}`; numeric-only labels and `GOTO`; reals
strictly `digits.digits[E...]` (no `1.`, no `.5`); consts are
`[sign] literal` (no expressions); enums without `= N`; set
elements without `..` ranges; untagged variants rejected
(`CASE tag: type` required); `TEXT`/`FILE` types absent
(`FILE OF` → 399 unimplemented); `PACK`/`UNPACK`/`PRED`/`SUCC`
parse but unconditionally mark 399 (stubs — kept out of the
battery); `NEW` exists but `DISPOSE` does not (`RELEASE` does);
no `ROUND`/`HALT`/`EXIT`/`RETURN`; no procedural types (the
compiler crashes on them); no `EXTERN`; `FORWARD` exists;
`WRITELN`/`READLN`/`READ`/`EOF`/etc. require parentheses
(`WRITELN;` is error 9) while user routines allow bare calls;
`EOLN`/`READ` take file args; `PACKED`, multi-dim arrays,
subrange/const bounds, `CHAR` indices, pointer forwards,
`OTHERWISE`-less everything.

## Grammar notes

Single-statement `if`/`while`/`for` branches (standard Pascal, avoids
`;`-capture across case items), `begin..end` compounds as statements,
trailing-`;`-tolerant sections/fields/case-items, order-free
repeatable declaration sections (as in pcom's `BLOCK`), mandatory
`(`...`)` on the 13+16 predeclared routines (`StdProc`/`StdFunc`,
which is what rejects bare `WRITELN;`), write-width params,
`PROGRAM name[(files)]` headings, strict EOF check (patched into
`Parse` by `build.sh`, which also fails on degenerate `IF/WHILE
FALSE` generation).

Lenient acceptances (documented misses, never false positives):
`DISPOSE`/`ROUND`/user-`TEXT` parse as ordinary identifier calls/
types although pcom marks them; lowercase is rejected while fpc
would accept (P2 direction wins — pcom is the oracle here).

## Build

```sh
./build.sh        # -> build/P2 (build/ is git-ignored)
```

Needs `CR` (Coco/R), `gm2 -fiso`, the `parser.frm`/`scanner.frm`
frames and `FileIO.def/.mod` — all resolved from the environment with
sensible defaults (`CRFRAMES`, `M2LIB`, `CR`, `GM2`).

## Differential results (P2 vs `pcom`)

Accept battery (23 cases: all statements, `NEW`/deref/`NIL`,
`ORD`/`CHR`, reals, strings, sets, `ODD`/`TRUNC`, trig, `EOF(INPUT)`)
+ `tests/deep.pas` (labels/goto, nested routines with `var` params,
records, enums, subranges, pointers, all statements): **0 false
positives**. Broken mutants (13, incl. spin-cases `&`, `{}`, `@`,
unterminated comment, `ELSE`, ident labels, `1.5.2`, bare
`WRITELN`): **0 missed**, positions close. Real P2 corpus
(`Pascal-P2-master/sample_programs`, `p2/Examples`): 5/5 agree
(`qsort` needed the trailing-newline normalization).