ltm: scalar feeder of a whole-RHS variable-backed reduce emits non-compiling link scores (warned 0-stub loops instead of the #758 drop) #790

Summary

The SCALAR feeder of a whole-RHS variable-backed reduce gets link-score variables from try_scalar_to_arrayed_link_scores whose fragments FAIL to compile, so every loop through the feeder degrades to a warned constant-0 stub -- instead of either being scored (the machinery exists for the subexpression spelling) or dropped via unscoreable_edges (the #758/#780 contract for unscoreable edges).

Fixture (reproduced at commit ef91194d with a TestProject fixture)

D1 = {a, b}, D2 = {c, d}
pop[D1]    -- stock
matrix[D1, D2]
scale      = SUM(pop[*]) * 0.005     -- scalar
growth[D1] = SUM(matrix[D1, *] * scale)   -- whole-RHS variable-backed reduce
... feedback closed through pop -> scale -> growth -> pop

The user equation compiles and simulates correctly. LTM emits $:ltm:link_score:scale->growth[a] / [b] via try_scalar_to_arrayed_link_scores, whose fragments fail to compile (Assembly warnings, constant-0 slots), and the scale-closure loop scores u1/u2 degrade to warned 0-stubs.

Why this is a gap, not just a degradation

• The subexpression spelling of the SAME dataflow scores fine: growth[D1] = 0.1 + SUM(matrix[D1,*] * scale) routes the scalar feeder through generate_scalar_feeder_to_agg_equation and yields a real score (~0.52). So the scoring machinery exists; the whole-RHS variable-backed-reduce target just never reaches it -- try_scalar_to_arrayed_link_scores does not handle that target shape.
• When a shape genuinely cannot be scored, the standing contract (ltm: declined element-mapped sliced reducer emits a non-compiling scalar link score on the conservative path (loop scores stub to 0) #758, and ltm: PartialEquationError link-score skips don't drop dependent loop scores via unscoreable_edges #780 for the remaining gap in it) is ONE loud warning + NO link-score variable + dependent loop scores DROPPED via unscoreable_edges. Emitting variables that are known not to compile produces the warned-0-stub middle state both of those issues exist to eliminate: constant-0 series that look like computed values, plus cascading per-loop warning noise.

So this is loud (warned), not silent -- but it is the wrong degradation mode AND an expressiveness gap with a working sibling path.

Components affected

• src/simlin-engine/src/db/ltm/link_scores.rs -- try_scalar_to_arrayed_link_scores does not handle the variable-backed-reduce target; the emitted fragments reference the reduce shape in a way that fails assembly. Fix should either route the scalar feeder of a variable-backed reduce through the agg-feeder path (generate_scalar_feeder_to_agg_equation, matching the subexpression spelling), or decline the shape into unscoreable_edges per the ltm: declined element-mapped sliced reducer emits a non-compiling scalar link score on the conservative path (loop scores stub to 0) #758 contract.

Related

• ltm: declined element-mapped sliced reducer emits a non-compiling scalar link score on the conservative path (loop scores stub to 0) #758 (closed) -- established the loud-skip + loop-drop contract this shape should follow if declined
• ltm: PartialEquationError link-score skips don't drop dependent loop scores via unscoreable_edges #780 -- the PartialEquationError leg of the same contract gap (skips that don't populate unscoreable_edges); this issue is the emit-then-fail-compile leg
• ltm: arrayed-owner Pinned/subset broadcast reduce slice loud-skipped; needs the PerElement broadcast rule on variable-backed owners #777 -- the arrayed-owner NO-Iterated-axis variable-backed sibling (declined loudly; distinct shape -- this fixture's slice HAS an Iterated D1 axis)
• ltm: loop scores through a variable-backed partial-reduce reducer reference never-emitted link-score names, stubbing to 0 #752 / ltm: synthetic agg node for SUM(arrayed[*] * scalar) with a scalar target fails to compile and is stubbed to constant 0 #738 (closed) -- earlier variable-backed-reduce and scalar-coefficient agg compile siblings
• Part of LTM tracking epic epic: Loops that Matter (LTM) -- tracking known issues and improvements #488 (cluster A)

Discovery context

Empirically confirmed during the adversarial review of commit ef91194d ("engine: decline bare-spelled feeder of un-hoisted reducer loudly", the GH #779 fix, branch ltm-phase1-residuals), reproduced at that commit with TestProject fixtures. Pre-existing, not a regression from that commit; out of scope for #779's fix.