ltm: FixedIndex reference into a disjoint-dim ApplyToAll target is loud-skipped though recoverable via the #510 per-element construction #769

Summary

The GH #758 loud-skip gate (commit 5c9c2d41 on branch ltm-fix-batch-2: skip dim-incompatible conservative link scores -- target_dims == [] with both endpoints arrayed non-module -- with one Warning, no link-score variable, and loops through the edge dropped) sweeps in a sub-case that is recoverable in principle: a FixedIndex reference into a disjoint-dimension ApplyToAll target.

Fixture shape:

pop[D1]                      -- source over D1; a1 is a literal element of D1
hub[D2] = pop[a1] * 0.05     -- ApplyToAll target over disjoint dimension D2
... a feedback loop closed through pop and hub

The gate loud-skips this edge: one Warning, no link-score variable, and every loop through it is dropped. That is a strict improvement over the pre-#758 behavior (a never-compiling scalar score whose dependents stubbed to warned zeros -- verified), so nothing regressed. But the shape does not need to be skipped at all.

Why it is recoverable

Every target slot genuinely reads the same pop[a1], so a per-element link-score construction -- an Equation::ApplyToAll over the target's dims holding pop[a1] live -- is well-defined for every slot. This is the GH #510 Ast::Arrayed per-element construction generalized to ApplyToAll targets. Unlike the element-mapped cross-product class the #758 gate primarily exists for, there is no wrong-slot-diagonal hazard: the literal index pins the source element uniformly across all target slots.

Why it matters

Severity: low. This is a capability gap behind a loud warning, not silent wrongness. But the shape -- literal-element cross-dimension references in feedback loops -- appears in MDL-imported models, and each occurrence currently drops every loop through the edge from LTM analysis.

Components affected

• src/simlin-engine/src/db/ltm/link_scores.rs -- emit_per_shape_link_scores and try_disjoint_dim_arrayed_link_scores (both present at 5c9c2d41)

Fix direction

Extend the #510 per-element construction in try_disjoint_dim_arrayed_link_scores to accept FixedIndex references into ApplyToAll (not just Equation::Arrayed) disjoint-dim targets: emit $⁚ltm⁚link_score⁚{from}[{elem}]→{to} as an Equation::ApplyToAll over the target's dims holding {from}[{elem}] live. Edges handled this way then no longer reach the loud-skip gate's sweep, shrinking the unscoreable-edge set (which also shrinks the #768 pathway/composite residual surface).

Discovery context

Identified during adversarial review of the GH #758 work on branch ltm-fix-batch-2 (commit 5c9c2d41).

Tracking / related

Part of LTM tracking epic #488 (Core algorithm group) and array-support umbrella #273. Related but distinct:

• ltm: declined element-mapped sliced reducer emits a non-compiling scalar link score on the conservative path (loop scores stub to 0) #758 -- the loud-skip gate whose sweep this sub-case sits inside (the gate itself is correct and an improvement; this issue is about narrowing it)
• ltm: degenerate link score for disjoint-dimension arrayed-to-arrayed edges with per-element target equations #510 (closed) -- the per-element Equation::Arrayed-target construction this would generalize
• ltm: sub-model pathway/composite scores through an unscoreable (loud-skipped) or agg-routed edge silently multiply to 0 #768 -- pathway/composite scores through loud-skipped edges silently zero (fixing this issue shrinks that surface but does not fix it)
• ltm: iterated+literal mixed subscript keeps the cross-product element expansion -- phantom cross-element loops carry silent nonzero scores (rescoped: hard compile failure fixed) #525 -- partially-iterated arrayed subscripts (different shape)