ltm: PREVIOUS-captured array-valued non-reducing builtin (RANK) scalarizes into ill-typed stubbed helpers, corrupting link scores #742

Summary

When LTM scores a link whose target equation contains an array-valued, NON-reducing builtin subtree (RANK demonstrated; VECTOR SORT ORDER / VECTOR ELM MAP / ALLOCATE are plausibly the same class), the ceteris-paribus partial wraps that subtree atomically in PREVIOUS(...), builtins_visitor::make_temp_arg captures the complex PREVIOUS argument into per-element scalar helper auxes, the helper equation is ill-typed in scalar context, the helper fragments fail to compile and are silently stubbed to constant 0, and the consuming link score is corrupted -- with no diagnostic anywhere.

Demonstrated repro

With arrayed pop[region] and a scored feedback-loop edge scale -> grow:

grow[region] = scale[region] * RANK(pop, 1)

• The ceteris-paribus partial for the scale -> grow link embeds PREVIOUS(rank(pop, 1)) (the RANK subtree contains no live scale reference, so wrap_non_matching_in_previous wraps it atomically).
• builtins_visitor::make_temp_arg captures the complex PREVIOUS arg into per-element scalar helper auxes (...⁚arg0⁚north / ...⁚arg0⁚south) whose equation is rank(pop, 1).
• That equation is ill-typed in a scalar helper: RANK is array-valued and non-reducing, Pass 1 of the builtins visitor never decomposes RANK args, and the dimension-bounds machinery cannot rescue it (verified identical behavior at multiple commits).
• The helper fragments fail to compile; assemble_module silently drops them (the silent drop itself is ltm: failed implicit-helper fragment compile emits no diagnostic (model_ltm_fragment_diagnostics skips model_ltm_implicit_var_info) #741); the helpers read constant 0 at runtime.
• The scale -> grow link score reads a corrupted -1000 instead of the true value. No Warning fires.

Root cause

Two interacting pieces:

1. src/simlin-engine/src/ltm_augment.rs -- is_array_reducer_name (~line 299) is a thin reader of ltm_agg::reducer_kind_from_name, which includes RANK. The wrap_non_matching_in_previous call site (~lines 639-652) uses it to decide "this subtree reduces to a scalar, so wrapping it atomically in PREVIOUS(...) is safe." That is true for SUM/MEAN/STDDEV/SIZE/MIN/MAX, but RANK does not reduce -- RANK(arr, n) is array-valued -- so the resulting PREVIOUS(<array-valued expr>) has no scalar meaning.
2. src/simlin-engine/src/builtins_visitor.rs -- make_temp_arg (~line 694) synthesizes the captured PREVIOUS argument as scalar helper auxes. For an array-valued non-reducing builtin subtree there is no per-element scalar projection it can emit (unlike the bare-arrayed-name case engine: bare arrayed name inside nested PREVIOUS() in an apply-to-all equation fails to compile #541 fixed via arrayed helper synthesis), so the helper lands ill-typed.

The failure then disappears into the model_ltm_implicit_var_info silent-stub path tracked separately as #741.

Relationship to existing issues

• engine: bare arrayed name inside nested PREVIOUS() in an apply-to-all equation fails to compile #541 (closed): the same class -- scalar helper capture of array-shaped PREVIOUS args -- for a bare arrayed name. Its fix (arrayed helper synthesis / subscripting the bare reference) does NOT cover an array-valued builtin application: there is no name to subscript, and Pass 1 never decomposes RANK args. Verified byte-identical broken behavior at commits bd4376e and 9187da4.
• ltm: failed implicit-helper fragment compile emits no diagnostic (model_ltm_fragment_diagnostics skips model_ltm_implicit_var_info) #741 (open): the missing diagnostic that lets this corrupt scores silently; fixing ltm: failed implicit-helper fragment compile emits no diagnostic (model_ltm_fragment_diagnostics skips model_ltm_implicit_var_info) #741 makes this loud but still wrong.
• 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 (open): cousin in the same wrap_non_matching_in_previous -> ill-typed-PREVIOUS-arg family, but for partially-iterated subscripts (a hard compile failure, not a silent stub).
• docs/tech-debt.md entry 27 covers RANK's delta-ratio fallback when RANK is the scored reducer itself -- a different path; this issue is about RANK appearing as a frozen (non-live) subtree inside someone else's scored partial.

Why it matters

Silent wrong link scores (and therefore loop scores / dominance attribution) for any model where an array-valued non-reducing builtin appears inside a scored equation on a loop edge. Severity low-medium: the shape is uncommon (no current corpus model hits it), but when hit the corruption is invisible to the user.

Components affected

• src/simlin-engine/src/ltm_augment.rs -- is_array_reducer_name (~299) and its wrap_non_matching_in_previous call site (~639-652): the atomic-PREVIOUS wrap is only sound for genuinely reducing builtins
• src/simlin-engine/src/builtins_visitor.rs -- make_temp_arg (~694): scalar capture of array-shaped PREVIOUS args
• src/simlin-engine/src/db/ltm/compile.rs -- the fragment compile path where the helpers fail

Possible approaches

• In ltm_augment.rs, distinguish reducing from array-valued builtins at the wrap site: exclude RANK (and audit VECTOR SORT ORDER / VECTOR ELM MAP / ALLOCATE) from the atomic-PREVIOUS shortcut, instead recursing into the args and PREVIOUS-wrapping the scalar-shaped leaves (subscripting array refs by the active A2A dimension first, mirroring the engine: bare arrayed name inside nested PREVIOUS() in an apply-to-all equation fails to compile #541 generator-side workaround), so every captured helper is genuinely scalar.
• Alternatively, extend make_temp_arg to synthesize an arrayed helper aux when the captured expression is array-shaped (the generalization of engine: bare arrayed name inside nested PREVIOUS() in an apply-to-all equation fails to compile #541's fix from bare names to arbitrary array-shaped expressions), then subscript the helper reference at the capture site.
• At minimum, once ltm: failed implicit-helper fragment compile emits no diagnostic (model_ltm_fragment_diagnostics skips model_ltm_implicit_var_info) #741 lands its diagnostic, ensure this shape produces the loud constant-0 Warning rather than nothing (and consider refusing to emit the partial with an unscoreable-edge warning instead of a corrupt score).

Discovery context

Identified during adversarial review on branch ltm-core-batch; pre-existing on main, NOT a regression of that branch's work (verified byte-identical at bd4376e and 9187da4). Part of LTM tracking epic #488.