Strong Slot Typing — Quality & Testing Report

For: Fable 5 design review Scope: schema-0.2 strong slot typing (remote-control design §2.2) + the codex-review fixes Date: 2026-07-02 Branch/commits: voice-capability-hierarchy → main — 2513a7d4, c6871ce3, cc84fa6a, 8d991170 (+ design docs 10ff8b52, a2d6d991)

1. What shipped

The slot_types facility was extended (not replaced) to make collect slots strongly typed, enforced at compile time (riffc L0-10) and runtime (validators):

Unified type: field — semantic names (enum, phone, …) never collide with JSON types (string, …), so one type: carries both; {semantic_type: …} long-form still works.
enum value domain — options (scalar or {value, label, aliases}), validated at load (non-empty), enforced at runtime (value ∈ options, alias→canonical).
Presentation — label (screen), optional, widget (override, restyle-only per OD-R5), and reask canonized as the per-slot voice prompt (OD-R7).
date implemented — extraction (_extract_calendar_dates) + IMPLEMENTED_TYPES, closing a pre-existing gap.
riffc L0-10 — an untyped collect slot is a hard build error.

2. Codex findings → fixes → evidence

Codex reviewed the design + first slice and found real gaps. All slice-level findings are closed with tests:

#	Sev	Finding	Fix	Test evidence
1	HIGH	runtime ignored unified `type:` → `{type: phone}` accepted garbage	resolve via `effective_semantic_type` at all 3 validation sites	`test_unified_type_enforced_at_runtime`, `test_unified_and_longform_agree_at_runtime`
2	MED	`{type: phone, semantic_type: email}` silently resolved by precedence	reject conflicting semantics at load	`test_conflicting_semantics_rejected` (3 forms)
3	HIGH	`enum` declared but not enforced at runtime	`enum_match` membership check in `validate_slot_value`	`test_enum_accepts_and_canonicalizes` (5), `test_enum_rejects_out_of_domain` (5), `test_enum_rejects_injection_before_membership`
7	HIGH	channel-bind unsound (design)	resolved by Fable 5 → §3.1 device-pairing (OD-R6)	design doc
8	MED	L0-11 undefined field-voice-prompt (design)	resolved → canonize `reask` (OD-R7)	design doc

Design findings #7/#8 are in Fable 5's court (recorded as OD-R6/OD-R7 §9). Surface-lane findings #4 (GoalSegment target_slots), #5 (enum option metadata), #6 (untyped-degrade) are assigned to the parallel surface session with the iter_slots(flow) directive.

3. Acceptance A5 coverage

A5: riffc rejects a surface button naming a nonexistent token, a form field with an untyped slot, and a screen-only state with no voice path.

Untyped-field leg — DONE & tested: L0-10 makes an untyped collect slot a hard error. test_l0_10_untyped_synthetic_flow_errors (error-level, names the slot), test_l0_10_typed_synthetic_flow_compiles_clean. Corpus: 3 golden anchors emit 0 L0-10 errors; migration checklist enumerates the remaining 10 flows.
Token + screen-only legs are in the parallel surface slice (L0-11/L0-12).

4. Test inventory

Suite	Tests	Focus
`test_slot_types_schema02.py`	16	vocabulary, validation, unified-type + enum runtime, L0-10
`test_slot_typing_adversarial.py`	53	adversarial: every semantic via `type:`; enum alias/case/whitespace/injection/collision; conflicts in every form; runtime enforcement per semantic; widget override; `date`; L0-10 synthetic flows
`test_typed_slots_across_flows.py`	157	cross-fleet: every typed slot resolves to an implemented type + validates type-correctly
`test_slot_type_ratchet.py`	2	exact ratchet — no untyped behavioral slot silently added
`test_riffc_golden.py` / `test_riffc_all_flows.py`	3 + 60	anchor census/checks locked; whole-corpus classification

Adversarial highlights (the "under pressure" evidence): enum rejects an injection payload ("drain\x00; DROP TABLE") via the shape check before membership; alias collisions resolve deterministically by declaration order; conflicting type:/semantic_type: rejected; {type: phone} and {semantic_type: phone} proven byte-identical at runtime.

5. Regression discipline — every remaining failure is pre-existing

The full -k "validator or slot" suite has 11 failures + 1 error. Every one was verified failing on clean HEAD (41792be4, before any typing work) via isolated git worktree runs — pizza/scheduling/offer/fsm-debug scenario tests unrelated to typing. This slice:

introduced zero new failures, and
net-fixed one pre-existing failure (apartment_intake:move_in_date=date) by implementing date (test_typed_slots_across_flows: was 1-failed/150-passed → now 157/157).

Method note: each suspected regression was reproduced on a clean checkout before attribution — no failure was assumed mine or pre-existing without proof.

6. Honest caveats / open items

main carries ~11 pre-existing test failures (pizza slot-extractor, austin acceptance/readback, scheduling offer loop) unrelated to typing — flagged, not owned by this slice.
enum extraction is validation + membership only; free-speech→option extraction (constrained decode for collect) is a later slice — declaring an enum today validates a filled value but relies on the LLM/other extractors to propose it.
Surface-lane #4/#5/#6 (GoalSegment, enum metadata, untyped-degrade) are open in the parallel session; my L0-10 should adopt iter_slots once it lands (single enumeration contract).
The I5 emission-location debt (state broadcast from transport, not engine bus) is separate and tracked in the console-map plan.

7. Verdict

Strong typing is real end-to-end: declared at load, enforced at compile (L0-10) and runtime (unified type: + enum membership), covered by 228 passing typing tests (schema02 16 + adversarial 53 + cross-fleet 157 + ratchet 2) including a 53-case adversarial suite, with a proven-zero-regression, net-positive impact on the corpus.