RTTI: field get/set by name (extends the VMT-8 method-reflection blob)

• Track: A (compiler core: RTTI blob layout, symtab, codegen).
• Opened: 2026-07-19, needed by [[feature-lib-pyexec]] (uforth arc) — and by the LFM/streaming plan (project_rtti_streaming_plan), dual use.

Method reflection exists (VMT-8 blob: names + invoke-by-name, shipped in the fpcunit arc). Missing: FIELD reflection — per-class table of (name, type kind, byte offset, record id for aggregates), plus runtime helpers:

• GetFieldPtr(obj, 'name') -> Pointer (+ type kind out) — get/set built on it.
• Works for Pascal classes AND NilPy-lowered classes (same class machinery; the exec bridge resolves vm.here / vm.memory[i] = x through this).

Notes:

• Blob names as frozen strings (same as the method blob — see project_rtti_reflection_and_overload_landmines).
• Layout addition rides the existing RTTI blob emitter; bump its version tag, keep old readers working (streaming plan reads the same blob).
• Self-host: compiler binary carries RTTI blobs -> codegen change = reseed/codegen-differ expected on the landing commit.

Gate: make test + self-host byte-identical; regression exercising field get/set by name on a class with mixed field types incl. a record field and a variant.

2026-07-21 finding — the current field RTTI is INSUFFICIENT for this

Measured while scoping the uforth exec() bridge (pyeval). Two concrete gaps the implementation MUST close, both verified in rtti_emit.inc:

• Published-gated. EmitRTTI gives unpublished classes "zero counts and no prop/meth arrays" (rtti_emit.inc ~329). NilPy classes publish nothing, so uforth's VM gets an EMPTY field table today — reflection finds nothing. The field/method tables must be emitted for ALL NilPy-lowered class members, not just published ones. (The per-class HEADER is already emitted for every class; only the member arrays are gated.)
• No field TYPE. EmitFieldInfo (rtti_emit.inc ~295) writes only name-ptr + byte offset. The exec bridge reads/writes vm.memory / vm.here / vm.base as VARIANTS, so it needs the field's type kind (and record id for aggregates) to box/unbox correctly — exactly the (name, type kind, byte offset, record id) tuple this ticket already specifies. Confirmed the gap is real, not hypothetical.

Consumer census (uforth PYTHON blocks): ~25 distinct vm members are touched — FIELDS memory, vars, here, rstack, stack, dict, base, _pic_buf, current_def_tokens, input_line, fstack, input_pos, current_token_index, current_def_name, xt_table and METHODS define_word, next_token_strict, next_token, run_forth_word, exec_token_runtime, strip_string_token, is_string_token, trace.

LANDED 2026-07-21 (commit e8ebbf0a) — field side complete

• Field table now emits EVERY field (not just published); FieldInfo grown to (name, offset, typeKind, recId, flags), RTTI_FIELD_SIZE 16 -> 40, flags bit0 = published so the streamer still filters. EmitFieldInfo/EmitRTTI in rtti_emit.inc, TFieldInfo in typinfo.pas kept in lockstep.
• Runtime helpers in typinfo.pas: GetFieldInfoByName, GetInstanceRTTI (VMT-8 backlink), GetFieldPtr(instance, cls, name, out kind) -> Pointer.
• test/test_rtti_field_get_by_name.pas: int/int64(read+write)/object/record + absent-name. Quick+limited GREEN (only the pre-existing test-core#122 flake), self-host byte-identical after the expected RTTI reseed.

FOLLOW-ON (not this ticket) — method table is ALSO published-gated

Discovered while landing fields: the METHOD table (UMthPub=1 gate in rtti_emit.inc) is published-only too, so GetMethodAddr finds NOTHING on a NilPy class — method invoke-by-name does NOT yet cover uforth's define_word / run_forth_word / etc. The exec bridge needs the same un-gating for methods, plus arity/param-type in MethInfo (currently name+code only) so the generic native-call trampoline can marshal args. Filed as part of the pyeval host-bridge work in [[feature-lib-pyexec]] (build with the trampoline), not here — this ticket was scoped to FIELD get/set and that is done.

Log

• 2026-07-21 — resolved, commit e8ebbf0a.