C shift-result-type battery (00200): result type = promoted LEFT operand across all int classes

• Type: bug (C type model — integer promotion + shift result type + sizeof-of-expression + unary-minus signedness). Track C.
• Split 2026-07-08 out of [[bug-c-expr-result-type-model]] (00178 + 00104 done; this is the remaining deep sub-effort).

Failing test

• 00200 (Vincent Lefevre's lshift-type.c): ISO C99 6.5.7p3 — for X << T, the integer promotions run on each operand and the RESULT TYPE is that of the promoted LEFT operand. The battery probes sign+size of the result across short/int/long/long long x signed/unsigned via PTYPE(M) = (sign of M via -(M)<0) * sizeof((M)+0). 30 cases fail.

Root cause

pxx mis-types shift results (e.g. (short)1 << (short)1 should be signed int — short promotes to int — but pxx types/treats it such that -(result) < 0 is false, i.e. unsigned). Needs: correct integer promotion (short/char -> int), shift result = promoted-left type (NOT the usual-arithmetic-conversions common type), sizeof-of-expression honoring that type, and signed unary minus keeping signedness. Interacts with the constant-type ladder just landed for 00104.

Gate

00200 exit 0 (30/30); drop from pxx.skip; c-conformance + corpus green; self-host byte-identical.

RESOLVED 2026-07-08 (fable-abc, Track A/C) — shift result = promoted LEFT operand

Root cause was narrower than the ticket feared: the sizeof-of-expression, unary-minus signedness, and integer-promotion pieces already worked. Only the SHIFT result type was wrong — CBinResultTk applied the usual-arithmetic- conversions over BOTH operands, so a wide/unsigned shift COUNT infected the result signedness: (short)1 << (unsigned long)1 came out unsigned (tyUInt64) instead of signed int. C99 6.5.7p3: the result type is the promoted LEFT operand; the right operand's type is irrelevant.

Fix (compiler/cparser.inc CBinResultTk): a shift branch (tkShl/tkShr) returning CIntegerPromoteTk(left) before the usual-arithmetic-conversion path (+ a forward decl for CIntegerPromoteTk). 00200 goes from 31 fails -> 0.

Gates (all green): 00200 "0 test(s) failed" (dropped from pxx.skip -> 205/0/15); regression test/cshift_result_type_b198.c in test-core; sqlite suite byte-identical; make test; self-host byte-identical; test-lua green.

Discovered alongside (independent, out of scope, filed): [[bug-c-sizeof-widening-cast-expr]] — sizeof((long)1) returns 4 not 8. 00200 only checks INTERNAL consistency (PTYPE(X)==PTYPE(X<<count)), so a uniformly under-sized long balances and does not surface it.

Log

• 2026-07-08 — resolved, commit ca169afa.