[paul-hammant]

Fixes #3479.

Reworks Jules's draft (paul-hammant#49, Jules task 7456359029880539017).

The bug

c = s[i] (a local, parameter, field or returned string) left c untyped. collect_expression_constraints set an element type for AST_ARRAY_ACCESS only when the base was TYPE_ARRAY, so a string base fell through, and codegen defaulted the local to int with warning: unresolved type in codegen, defaulting to int. That fails make check-examples, which blocks the Rosetta Caesar cipher in #3477. return c as int on such a local was refused too ("cannot cast unknown to int").

The fix (Jules's, 4 lines)

A string base types the read as byte, which is what the language reference says s[i] reads.

This changes behaviour, and it is consistent with byte[]:

• a high byte reads 195, not -61;
• arithmetic stored back into the local wraps: c = s[0]; c = c - 100 gives 244 for "X", where the int default gave -12. A local bound to a byte[] element (x = b[0]; x = x - 70) already wraps the same way on main;
• c = s[0]; c = 1000 is now a compile error (byte literal out of range);
• s[0] + 1, s[0] == 88 and match c { 65..=90 -> ... } are unchanged. s[i] as int keeps a signed result.

The changelog says so, and the language reference now says s[i] reads a byte.

What changed from the draft

• The test is an integration directory, not a regression file. The draft's tests/regression/test_issue3479_string_index_local.ae re-ran ./build/ae build --with=all from inside itself. That depended on the current directory, wrote into build/, and used --with=all, the backdoor AGENTS.md rules out. tests/integration/string_index_local_type/ has a fixture and a script. The script builds the fixture with no warning at all and runs it, and the sweep runs the fixture too.
• The fixture pins the byte semantics: 88 - 100 wraps to 244, and as int gives -12.
• The changelog says what the byte type means for arithmetic.

Verification

• Negative control 1: the fixture built with main's compiler fails on return c as int ("cannot cast unknown to int").
• Negative control 2: a copy without the casts fails the no-warning check, with four unresolved type in codegen, defaulting to int warnings.
• With the fix: [PASS] string_index_local_type.
• make test: passes.
• make test-ae: 2071 passed, 2 failed. The two are http_server_h2 and windows_crt_symbols, which fail the same way on a clean main.
• check-examples: clean apart from an untracked local directory.
• Gates: ae fmt, check-docs, check-changelog, check-tests, check-warnings and check-raw-calls all pass.

🤖 Generated with Claude Code

[paul-hammant]

Superseded by 235ea27 (in 0.871.0), which fixed #3479 on main with s[i] typed as an int from 0 to 255 rather than the byte this PR uses (a byte local wraps on arithmetic and refused uint32 u = s[i]). Closing.