RTTI is emitted unconditionally (every class, even a classless program) — dead weight on ESP32/embedded

• Track: A (RTTI emission is core: compiler/rtti_emit.inc, the emit driver in parser.inc). Tag: O (size/codegen). Cross-cutting toward the ESP/xtensa + riscv32 targets.
• Found: 2026-07-16, dissecting the frozen hello demo's size (280 B remembered vs 785 B now). See [[project_rtti_streaming_plan]], [[project_rtti_reflection_and_overload_landmines]].

What we measured (frozen, -uPXX_MANAGED_STRING)

Breakdown of the x86-64 baseline growth vs the old ~280 B:

• Signal-handler runtime ~272 B of code — x86-64 default-on, --no-signals opts out. NOT on ESP (x86-64-gated). Working as intended; leave it.
• Div-by-zero abort stub + "Runtime error 200…" string (~40 B) — x86-64 ONLY (if TargetArch = TARGET_X86_64 then EmitDiv0Stub, parser.inc ~24210). NOT on ESP (uses hardware/trap behavior). Fine.
• RTTI — emitted on ALL targets, including ESP. This is the only dead weight on a micro. The bare xtensa binary is code=6 B but still carries a TObject remnant in .data.

The two RTTI wastes

1. Every user class gets a full RTTI header even when never reflected. Deliberate today — EmitRTTI (rtti_emit.inc ~319-338) emits one RTTI_CLS_SIZE header + interned name per class so ClassName answers for any class (the published-only gate used to make ClassName a coin flip). Correct, but on ESP32 with classes and no is/as/ClassName/TypeInfo/streaming it is pure size cost.
2. A classless program still emits a TObject remnant (~8 B interned name). The "plain wrong" case: zero classes, zero reflection, yet RTTI bytes ship.

Goal

Opt OUT of RTTI to compress embedded size, keep it available (opt-in / on-demand) for debugging, reflection, and advanced software (streaming, fpcunit, dynamic dispatch).

Approaches (ranked)

1. Usage-driven emission (north-star). Emit a class's RTTI only if the program actually reflects on it — ClassName/ClassType/InheritsFrom, is/as against it, TypeInfo, published-member access, LFM streaming. Track a per-class "RTTI-touched" bit during parse/lower; EmitRTTI skips untouched classes. Zero cost when unused, no flag, no silent breakage. Hard part: cases the compiler can't see statically (LFM FindClass-by-name, dynamic streaming) need a force-on escape.
2. --rtti=auto|full|none flag (pragmatic first step). auto = usage-driven (or, interim, the current all-classes behavior); full forces every class (for dynamic/streaming reflection the compiler can't prove); none hard-off — a reflection op under none is a compile-time error, never a silent wrong value. ESP/bare default leans (auto, and skip the classless TObject remnant).
3. Minimum quick win, independent of the above: when UClsCount = 0 and no reflection op appears, emit NO RTTI at all (drop the TObject remnant).

Guard rails

• Never silently break reflection. is/as, ClassName, LFM streaming ([[project_rtti_streaming_plan]]), fpcunit ([[project_fpcunit_green_metaclass_self]]) all consume RTTI. Gating must be usage-aware or explicit; a stripped build that hits a reflection op must error at compile time, not miscompute.
• ESP already skips signals/div0 by target; RTTI gating should follow the same "lean-by-default on embedded, full on host" instinct.

Not in scope (separate, minor)

• x86-64 EmitDiv0Stub is unconditional even for division-free programs (~40 B). Could gate on "a div/mod site exists." Small host-only win; file separately if wanted.

• 2026-07-18 night (fable-O): claimed, reconned, PARKED on a design fork — EmitRTTI also emits FUNCTIONAL data (class-field finalize layouts at [VMT-16], ClassName backlink at [VMT-8]), so --rtti=none semantics need a user call. Filed [[decide-rtti-none-semantics]] (options + recommendation). Unclaimed; back to backlog until decided.

Elevated 2026-07-20 — this IS the --rtti=none answer

[[decide-rtti-none-semantics]] resolved as option C: usage-driven emission, and no --rtti=none flag ships at all. So this ticket is no longer an optimization sitting alongside a user-facing switch — it is the entire mechanism by which RTTI shrinks, and the ESP/size-constrained audience has nothing else to reach for. Any remaining text describing --rtti=none as a planned switch is superseded.