ltm: loop scores through a variable-backed partial-reduce reducer reference never-emitted link-score names, stubbing to 0

[bpowers]

Summary

A loop score through a variable-backed partial-reduce reducer (is_synthetic == false -- the variable itself is the agg) references link-score names that are never emitted, so the loop-score fragments fail fragment compile (Assembly Warnings) and are silently stubbed to constant 0.

Fixture shape

D1 = {a, b}, D2 = {x, y}
matrix[D1,D2] = stock[D1] * 0.1
stock[D1]     (fed by inflow)
inflow[D1] = SUM(matrix[D1,*])     -- the WHOLE RHS: a variable-backed agg

A feedback loop closed through the D1 stock that broadcasts into the D1 x D2 matrix.

Mechanism

1. Variable-backed aggs do not get read-slice-driven element edges -- per the src/simlin-engine/src/ltm_agg.rs module docs, the element-graph reroute and the link-score emitter both filter to is_synthetic aggs, and "Variable-backed aggs still take the conservative cross-product in the element graph." So the element graph contains phantom matrix[a,x] -> inflow[b] cross-element edges.
2. The cross-element loop-score equations built over those edges reference "$⁚ltm⁚link_score⁚matrix[a,x]→inflow"[b] (subscripted per-element form) and the bare A2A "$⁚ltm⁚link_score⁚matrix→inflow".
3. But the only link scores actually emitted for a variable-backed partial reduce are the per-(row, slot) scalars matrix[a,x]→inflow[a] from try_cross_dimensional_link_scores (src/simlin-engine/src/ltm_augment.rs).
4. The referencing loop-score fragments therefore fail to compile, assemble_module drops them, and each loop score keeps its layout slot reading constant 0 -- visible only as Assembly Warnings in diagnostics.

Verified pre-existing: identical behavior at the parent of the #528 fix on branch ltm-core-batch (commit bce4aa4c); not a regression of that branch's work.

Why it matters

Medium. Silent zero loop scores for cross-element loops through a whole-RHS partial reducer -- row_total[D1] = SUM(matrix[D1,*]) as an entire equation is a common Vensim shape. The model compiles and simulates; the degradation is only visible if the user inspects diagnostics.

Components affected

• src/simlin-engine/src/ltm_agg.rs -- variable-backed aggs (is_synthetic == false) excluded from the read-slice-driven element-graph reroute, leaving the conservative cross-product in place
• src/simlin-engine/src/db/ltm/loops.rs / loop-score equation construction -- composes link-score names (bare A2A + subscripted per-element) that the partial-reduce emitter never produces
• src/simlin-engine/src/ltm_augment.rs -- try_cross_dimensional_link_scores emits only the per-(row, slot) scalar forms

Possible approaches

• Give variable-backed partial reducers read-slice-driven element edges like synthetic aggs (retiring the conservative cross-product for them), so the phantom cross-element edges -- and the loop-score equations over them -- never arise; or
• Teach the loop-score name resolution the per-(row, slot) scalar forms that try_cross_dimensional_link_scores actually emits, so the cross-element loop-score equations reference names that exist.

Existing test coverage

build_model_with_failing_ltm_fragment in src/simlin-engine/src/db/ltm_unified_tests.rs is exactly this fixture -- it is the positive fixture for the fragment-diagnostics tests (the #547 test-debt subject), deliberately leaning on this genuinely-unfixed bug. Its doc comment (rewritten in bce4aa4c; it previously mis-attributed the failure to #528) describes this gap accurately. Whoever fixes this issue must replace that fixture (or land the #547-style injection hook) so the diagnostics tests keep a real positive case.

Related

• ltm: Piece 3 fragment-diagnostics positive tests are coupled to the lifetime of real fragment-compile bugs #547 -- the test-debt coupling: the diagnostics positive fixture's lifetime is tied to this bug
• ltm: agg→target link score over-subscripts an arrayed synthetic agg in the broadcast case (agg[D1] feeding to[D1,D2]) #528 -- the synthetic-agg broadcast over-subscription this fixture originally leaned on (now fixed); distinct -- this issue is the variable-backed name-resolution gap
• ltm: un-hoisted multi-source iterated-dim-feeder reducer emits silent garbage arrayed loop score (combined_read_slice slice-disagreement carve-out) #743 -- un-hoisted multi-source reducer garbage scores (different mechanism: combined_read_slice declines the hoist; here no hoist is involved at all)
• ltm: loop-score builder's PureScalar fast path never routes a loop through a scalar synthetic agg node (silent zero loop score) #737 / ltm: synthetic agg node for SUM(arrayed[*] * scalar) with a scalar target fails to compile and is stubbed to constant 0 #738 -- synthetic-agg routing / compile gaps (this issue is the variable-backed sibling class)
• LTM needs to support arrays #273 -- LTM array support umbrella

Discovery context

Root-caused during GH #528 work on branch ltm-core-batch (commit bce4aa4c), while disentangling why build_model_with_failing_ltm_fragment still fails after the #528 fix.