Handover: consumer-side wishlist for the loco ABI (lance-graph → ogar-loco session)

.claude/handovers/2026-08-05-1430-lance-graph-orchestrator-to-ogar-loco.md

@@ -0,0 +1,102 @@
+# Handover: lance-graph orchestrator → ogar-loco session (2026-08-05)
+
+> APPEND-ONLY. Consumer-side wishlist for the loco ABI, written after a
+> source-verified review of merged #241 (R4 data-first vocabularies, R5
+> statement boundaries) from the seat that will consume it: lance-graph's
+> rig/rs-graph-llm oracle loop and the compiled-template stack. Every item
+> respects #241's recorded-not-minted list — nothing here asks for gated
+> mints, and all ids stay the operator's.
+
+## What I did (context for this wishlist)
+
+- Verified #241 claim-by-claim against merged source (`0d39025`). All
+  load-bearing claims code-true: unforgeable shared half in
+  `VocabularyTable::compose` (test `the_composed_tables_shared_half_is_
+  unforgeable`), table-backed `CheckedVocabulary`, the four-way refusal in
+  `statement_bounds`, `DOMAIN_FLOOR = 0x90` with the compile-time pin.
+- Sketched the consumer loop this wishlist serves: LLM (via lance-graph's
+  shipped rig CompletionModel adapter) emits N candidate shortcode bodies →
+  `validate` + `statement_bounds` refuse the garbage → the surviving bag is
+  executed and scored by shipped falsifiers (the Strict/Aware projection
+  machinery). The refuse-don't-guess design of #241 is the LLM-output
+  firewall; `table()`'s doc already names oracle schemas as a consumer.
+
+## FINDING (one, small, letter-vs-spirit)
+
+**F-1 — validate() samples the domain hooks twice.** `check(&v)` runs, THEN
+`compose(&v)` samples the hooks again. For a hooks-unstable vocabulary the
+stored table is a second, unvalidated sampling: `check` can pass on answers A
+while the table freezes answers B — including a `body_refs` beyond 3, which
+re-opens the traversal edge the wrapper's proof claims closed
+(`program.rs`'s `debug_assert`s already half-know this: they name "a
+hypothetical non-deterministic vocabulary" and degrade to a safe index panic
+in release). Shared half unaffected (stamped from core); stable vocabularies
+unaffected. Severity: pathological-input-only, fail-loud not corrupting.
+
+## The wishlist (consumer-priority order)
+
+**W-1 — Compose-then-check hardening (closes F-1).** Compose the table
+first, then validate: keep the method-drift half of `check` exactly as is
+(drift detection inherently needs the method reads), and add a shape check
+over the STORED table's own `body_refs` column so the artifact everything
+reads is the proven object. A few lines + one can-fire test with an
+unstable-hooks vocabulary. Makes "what was validated is literally what gets
+read" true to zero resamplings, before the second vocabulary lands on top.
+
+**W-2 — The name column (the prompt-legend enabler).** The table carries
+semantics but no names: `FnIndex` constants are named in code only. An LLM
+legend / oracle schema / renderer needs `name(f) -> Option<&'static str>` —
+shared-core names in the core, a `domain_name` hook for siblings, composed
+into the table (or beside it) under the same unforgeable-shared-half rule.
+Without it every consumer hand-rolls a name map that can drift; with it the
+legend is a serialization of the validated table itself, at the membrane
+where serialization is legal.
+
+**W-3 — The NARS-34 vocabulary mint (ids operator-gated, as always).** The
+34 rung-3 tactic recipes as the first reasoning vocabulary above the floor.
+Two constraints discovered from the consumer side:
+- The slot arithmetic forces recipes-not-atoms anyway: the domain range
+  holds 112 slots; the 144 rung-2 atoms don't fit one sibling, and their
+  unification is gated. 34 fits with room.
+- `pushes_result` must be declared per verb at mint time, and the mint is
+  where the `ResultBehavior`/`statement_terminal` question comes due: a
+  declared-pushing verb whose result is legally discarded produces SILENT
+  MASK-GRANULARITY COARSENING, not refusal — depth never returns to zero
+  mid-body, and the trailing-value rule merges intended statements into one
+  mask unit. Execution stays correct; per-statement masking silently
+  degrades. The can-fire test for that trace should ride in WITH the mint.
+
+**W-4 — Funnel telemetry as data (the oracle contract's safe half).** The
+generate-and-filter loop needs refusal STATISTICS: per candidate, which gate
+killed it (`ConformanceError` / `StatementError` variants are already typed
+— a small summary struct over a batch suffices). Distinction worth pinning
+in the doc when this lands: **validity feedback** (your candidate refused at
+underflow, call 7) is safe to loop back to the generator raw; **fitness
+feedback** (your candidate scored κ=X) falls under lance-graph's
+observer-effect payload law (distribution shape × rank, never the raw
+scalar) and stays out of the loop until that instrument exists. #241 parked
+the oracle contract; this is its smallest safe slice.
+
+**W-5 — The >64-statement split contract (doc-only).** `statement_bounds`
+already yields the count; what a lowerer must DO at statement 65 is
+undocumented ("a split signal, not a mask-widening use case" — but split
+how: second function? second template?). One paragraph in the module doc
+prevents N divergent conventions.
+
+## Explicitly NOT asked
+
+No 144-atom unification (gated). No structured PA control mints (ids the
+operator's). No replay/session representation. No cross-repo dependency in
+either direction — the consumer wants the TABLE AS DATA (W-2 makes it
+self-describing); lance-graph never becomes an ogar-loco dependency and
+ogar-loco never imports an LLM adapter.
+
+## Open questions for the loco session
+
+- OQ-1: Should the name column live IN `FnSpec` (one more field, table stays
+  the single artifact) or beside it (keeps `FnSpec` `Copy`-tiny)? Consumer
+  has no preference beyond "same unforgeable-shared-half rule".
+- OQ-2: `DOMAIN_FLOOR = 0x90` makes the shared range exactly 144 slots — the
+  same number as the rung-2 verb atoms. Almost certainly coincidence (the
+  core covers ~50 today), flagged only so the option is visibly open if the
+  gated unification question ever returns. Numerology until ruled otherwise.