ltm: de-hoisted array-valued RANK gets diagonal-only element edges, so rank-mediated cross-element loops are not enumerated #776

Summary

The #771 fix (de-hoist: reducer_is_hoistable now requires reducer_collapses_to_scalar, invariant I5 of docs/design-plans/2026-06-11-ltm-shape-expressiveness.md, commit 0b14a9e3 on branch ltm-shape-phase1) routes array-valued RANK references onto the Direct conservative paths. That kills #771's defect (the cannot-compile scalar synthetic agg and its zeroed agg-routed loop scores), but the resulting element-edge treatment is spelling-dependent, and for the bare-arg spelling it under-represents RANK's true dataflow: RANK(pop, 1) read at element e is an order statistic that depends on every element of pop, not just pop[e].

Mechanism: two spellings, two arms

The de-hoisted classification sends the two legal spellings of the same RANK read down different conservative arms (established by the T1 adversarial review of commit 0b14a9e3):

• Bare-arg RANK(pop, 1) -- classifies as a Direct/Bare site and takes the diagonal arm, emitting only pop[e] -> grow[e] element edges. Off-diagonal rank-mediated influence is missing, so cross-element loops through the rank reordering are not enumerated. This is the under-enumeration this issue tracks.
• Wildcard-arg RANK(pop[*], 1) -- yields a Direct Wildcard site with in_reducer=false (the ltm: sliced reducer sub-expressions are not hoisted into aggregate nodes (stay on the wildcard-link-score path) #514 reclassification doesn't fire), which takes the conservative cross-product arm. Rank-mediated cross-element loops are enumerated there -- reviewer-demonstrated: pop[north] -> {grow[north], grow[south]}, loops enumerated, scores ~0 under constant ranks, zero warnings. This spelling is safe but noisy (conservative over-enumeration).

So the residual is spelling-dependent: bare-arg = diagonal under-enumeration (the filed concern); wildcard-arg = conservative over-enumeration (safe, noisy). Two spellings of the same dataflow get inconsistent loop sets.

What is lost (bare-arg spelling)

Cross-element feedback loops mediated by rank reordering are not enumerated. Concretely: if pop[b] overtaking pop[a] changes the rank assigned at a and therefore what flows to a's consumer, that b -> a influence has no off-diagonal element edge, so the loop through it never appears in the element graph and is never scored.

Current residual behavior (loud-ish, conservative)

• Bare-arg: diagonal element edges only; same-element loops through RANK are enumerated and scored.
• Wildcard-arg: full cross-product edges; rank-mediated cross-element loops are enumerated, with zero warnings.
• For genuinely-constant rank orderings the rank-mediated link contributes ~0 score on either arm (RANK output is piecewise-constant; its delta-ratio stand-in is ~0 between reorderings -- cf. docs/tech-debt.md entry 27's RANK delta-ratio history).
• Pinned by rank_frozen_subtree_link_score_scores_correctly (src/simlin-engine/tests/integration/ltm_array_agg.rs), whose doc comment describes this residual; also documented in the reducer_collapses_to_scalar rustdoc (src/simlin-engine/src/ltm_agg.rs).

So this is a documented under-enumeration (missing loops) on the bare-arg spelling, not silent wrong numbers on enumerated loops -- consistent with the epic's "no silent wrong numbers" invariant, but short of full expressiveness, and inconsistent across spellings.

Why it matters

Low. RANK-mediated feedback is rare (no corpus model exercises it), and RANK's partial is non-differentiable at reordering points, so even an enumerated rank-mediated loop would score via a conservative stand-in. But a model whose dynamics genuinely hinge on rank reordering inside a loop would have that loop invisible to dominance analysis if spelled bare-arg -- and the same model gains or loses loops by rewriting RANK(pop, 1) as RANK(pop[*], 1).

Components affected

• src/simlin-engine/src/ltm_agg.rs -- reducer_collapses_to_scalar, the de-hoisted classification path
• the element-graph builder's edge emission: the Direct/Bare diagonal arm and the Direct/Wildcard cross-product arm (the ltm: sliced reducer sub-expressions are not hoisted into aggregate nodes (stay on the wildcard-link-score path) #514 reclassification gate)
• src/simlin-engine/tests/integration/ltm_array_agg.rs -- rank_frozen_subtree_link_score_scores_correctly pins the residual

Possible approaches

1. Arrayed agg node for array-valued reducers (the shape the de-hoist deliberately deferred, and ltm: enumerate_agg_nodes hoists whole-extent array-valued RANK into a scalar agg node that cannot compile (agg-routed loop scores zeroed) #771's noted alternative): mint a synthetic agg node whose result_dims are the argument's dims, with conservative all-to-all source->agg element edges. Enumerates the cross-element loops AND unifies both spellings' treatment (one arm regardless of bare vs wildcard arg); scoring stays conservative (delta-ratio).
2. Accept as documented limitation: RANK in feedback loops is rare, its partial is non-differentiable, and the current behavior is conservative and pinned. Keep this issue as the record and close as wontfix-by-design if the cost/benefit doesn't change -- though the spelling-dependent inconsistency weakens this option relative to the original filing.

Discovery context

Identified during T1 of the LTM shape-expressiveness Phase 1 work (branch ltm-shape-phase1, commit 0b14a9e3), as the deliberate residual of the #771 de-hoist. The spelling-dependence of the residual (bare-arg diagonal vs wildcard-arg cross-product) was established by the T1 adversarial review of the same commit. Part of LTM tracking epic #488 (cluster A). Related: #771 (the de-hoisted defect), #742 (closed capture sibling), docs/tech-debt.md entry 27 (RANK delta-ratio partial, by-design).