ltm: bare-spelled feeder of un-hoisted multi-source reducer gets silent wrong changed-last loop score (per-element partial vs broadcast execution) #779

Summary

The BARE-spelled feeder of an un-hoisted multi-source reducer gets a silently wrong (unwarned, sustained) loop score from the GH #743 changed-last partial: the partial compiles per-element while the executed equation broadcasts the bare reference.

Fixture:

D1 = {a, b}, D2 = {c, d}
matrix[D1, D2]                            -- co-source
frac                                      -- feeder, referenced BARE (unsubscripted)
growth[D1] = SUM(matrix[D1, *] * frac)    -- execution iterates the bare frac over D1
... feedback loops closed through matrix and through frac

This is the unfinished bare-spelling sibling of #743, whose fix covered the SUBSCRIPTED feeder spelling frac[D1] (now fully hoisted by the T5 feeder-sub-slice acceptance, #767).

Mechanism

1. The bare frac reference classifies all-Reduced -- the read-slice vocabulary cannot express a bare iterated read (a reference that execution iterates over the owner's A2A dimension without a subscript). Verified byte-identical pre/post T5 at commit bcb89565 on branch ltm-shape-phase1: this is PRE-EXISTING, not a T5 regression.
2. Because the slices disagree, the reducer stays un-hoisted. The matrix-closure loops land on the LOUD degraded path (5 fragment warnings + 0-stubs) -- that half is fine.
3. The frac-closure loop falls to the ltm: un-hoisted multi-source iterated-dim-feeder reducer emits silent garbage arrayed loop score (combined_read_slice slice-disagreement carve-out) #743 changed-last fallback partial, sum(matrix[d1,*] * PREVIOUS(frac)), which compiles per-element -- while the executed growth equation broadcasts the bare frac across all D1 rows (in the probe, growth_0 = 10 = the sum of all 2x2 terms).
4. Result: a sustained, UNWARNED loop score of ~2.92 for the frac-closure loop, internally inconsistent with the four co-source loop scores of +1 each.

Why it matters

Severity: low-medium. This is a silent wrong number (~3x) for a natural spelling of a common shape (per-row total = SUM(row * shared coefficient) with the coefficient written bare), violating epic #488's standing "no silent wrong numbers" invariant. The subscripted spelling frac[D1] is now fully correct after T5/#767, so users hitting this have an easy workaround they don't know they need.

Components affected

• src/simlin-engine/src/ltm_agg.rs -- read-slice classification (compute_read_slice / combined_read_slice): no vocabulary for a bare iterated read
• src/simlin-engine/src/ltm_augment.rs -- the changed-last chooser accepts the bare-feeder shape and emits a per-element partial that disagrees with the executed broadcast
• src/simlin-engine/src/db/ltm/link_scores.rs -- the emitted frac-closure loop score

Possible approaches

• Teach the slice vocabulary a bare iterated read (likely the right fix): classify the bare feeder reference the way execution treats it -- iterated over the owner's A2A dim -- so T5's feeder-sub-slice acceptance covers it and the reducer hoists exactly like the frac[D1] spelling. Cf. how classify_iterated_dim_shape already treats bare refs as Bare/diagonal.
• Or have the changed-last chooser decline the bare-feeder shape loudly (warning + 0-stub), restoring the no-silent-wrong-numbers invariant without the hoist.

Discovery context

Demonstrated by a reviewer during T5 review of the LTM shape-expressiveness phase (branch ltm-shape-phase1, commit bcb89565). Verified byte-identical pre/post T5 -- pre-existing, out of scope for that phase.

Tracking

Part of LTM tracking epic #488 (cluster A) and the array-support umbrella #273. Related: #743 (the subscripted-spelling sibling, fixed -- its changed-last fallback is the partial that goes wrong here), #767 (the feeder-sub-slice hoist extension that fixed the subscripted spelling's co-source closure, closing via the shape-expressiveness phase PR).