Theia-Modula-2/grammars/Blaise

Blaise Pascal validator pilot

blaise_pascal.atg is a genuine Blaise Pascal grammar authored from Blaise/blaise-master/docs/grammar.ebnf (Phases 1-9a) — not the Blaise-grammar/blaise_pascal.atg installed next to the IDE sources, which is byte-identical to the classic-Oberon grammar apart from its tag line and describes Oberon, not Pascal.

The EBNF turned out to be an incomplete guide to the language, so the grammar was developed against the blaise v0.14.0-SNAPSHOT oracle binary instead (~80 oracle probes settling every uncertain point), plus a 403-file project corpus. Findings (all encoded in the .atg header comment):

• The EBNF expression section omits or, in, unary minus, #nn character literals and float literals — the oracle accepts all.
• The EBNF statement section omits break, continue, case..of..(else)..end and repeat..until — all oracle-proven.
• Interface and implementation sections routinely carry type and var sections (300+ corpus files) — the EBNF allows none.
• Declaration sections come in any order and repeat (const before type, ...); statement lists and routine bodies tolerate a missing ; before the closing keyword.
• set of (A, B, C) (inline enum) is EBNF-documented but oracle-rejected; dotted names in const/array/set positions are oracle-rejected too (both accepted here — lenient superset).
• TBox(T) = class (paren form) is oracle-rejected; only TBox<T> declares a generic type. class(TObject); (forward, no end) is oracle-accepted although the EBNF requires end.
• Bare Foo; calls, inherited; without parens, := in var sections and =-style defaults are all oracle-rejected with explicit messages; the grammar matches.
• Contextual keywords (static, read/write, property, reference, on, Supports, ...) were each probed against the oracle and either kept as identifiers with structural handling or reserved — see the header comment.

Known validator gaps (all zero-impact on the corpus unless noted):

• Explicit generic free-function calls f<T>(x) (zero corpus use).
• Typed-parameter arrow lambdas (zero corpus use).
• asm routine bodies (RTL files only, documented).
• Nested procedures inside methods (rare, documented).

TEXT_BLOCK literals (''' ... ''') cannot be expressed in Coco/R tokens (the merged scanner DFA cannot tell '''-open apart from string-escape), so a small length-preserving pre-scan filter (TextFilter.mod) rewrites blocks to plain strings before parsing. Positions are preserved exactly; the shipped BlaiseV is a front-end script (BlaiseParse + BlaiseTF + temp-file handling).

Build

./build.sh        # -> build/BlaiseV (+BlaiseParse, BlaiseTF; build/ 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). Note: link via the two-step module list (-fgen-module-list then -fuse-list); a single-step link trips over pre-existing errors in this gm2 build's own ISO library .def files (RndFile/TermFile/ChanConsts set-constructor complaints).

Coco/R reports 0 errors and only benign LL(1) warnings (loop-exit and documented first-alternative orderings, each triaged).

Differential results (BlaiseV vs blaise)

• Accept battery (statements, expressions, types, classes, generics, lambdas, exceptions, char/textblock literals): 0 false positives.
• Broken mutants: 0 missed, positions close.
• tests/deep.pas: compiles clean with the oracle AND validates clean (const/type/var/threadvar sections, records, classes with static members/properties, interfaces, procedural/reference types, generics, all statements, lambdas, exceptions).
• Full project corpus (403 files): 387 clean-agree, 0 missed; the 16 remaining validator-FPs all contain asm routine bodies (documented gap above), and the 5 oracle-rejected files are correctly rejected by the validator too.