ltm: residual ~30k synthesized PREVIOUS helper-aux vars on C-LEARN threaten the VM 65k-slot ceiling #654

Summary

LTM instrumentation on C-LEARN v77 (test/xmutil_test_models/C-LEARN v77 for Vensim.mdl) with LTM discovery mode still synthesizes ~30k helper-aux variables, down from ~145k after the work in commits 7e6048e7, 82d20304, and 51392894 on main. C-LEARN now fits under the bytecode VM's 65,536-slot limit (56,891 slots) and works end-to-end, but the remaining helpers cost compile time, memory, and runtime, and a moderately larger model would hit the slot limit again.

This is the residual scalability/cost half of the PREVIOUS-helper-synthesis problem: it is distinct from #587 (a subscripted-PREVIOUS LTM fragment silently dropping its score to 0 -- correctness/accuracy) and from #647 (the discovery algorithm being intractable at element-expanded scale). This issue is specifically about the volume of helper-aux variables the augmentation emits, the slot/compile-time/memory pressure that creates, and how close it sits to the hard VM slot ceiling.

Remaining helper sources (measured on C-LEARN v77, LTM discovery mode)

1. ~9.7k PREVIOUS calls with bare-element subscripts that cannot be qualified. Two sub-cases: (a) elements declared by multiple dimensions at different positions (genuinely ambiguous), and (b) subscripted references of variables whose dims are not in scope at the wrap site. Root cause: wrap_index_non_matching_in_previous / qualify_element_index in src/simlin-engine/src/ltm_augment.rs (~:737-:766) only have the project DimensionsContext, not the subscripted variable's declared dims, so they cannot position-match the element and fall back to synthesizing a helper.
2. ~2k module-output refs (PREVIOUS(module·output)). These have fixed slots too and could compile to a direct LoadPrev with submodel-offset resolution instead of a synthesized helper.
3. ~2k index-wrapped refs (x[PREVIOUS(elem)]) from paths where qualification falls back due to ambiguous element membership -- same root cause as (1).
4. The per-target-element generators (~9.4k helpers) plus remaining lowercase previous(...) calls embedded from original equations.

Why it matters

• Headroom against the hard slot ceiling. C-LEARN now compiles at 56,891 slots vs. the 65,536-slot VM limit. The ~30k residual helpers are a large fraction of that budget; a moderately larger arrayed model that auto-flips to discovery would blow the ceiling and fail to compile. This is a structural scalability limit in the LTM-augmentation path, not just a polish item.
• Compile time, memory, runtime. Every helper is an extra synthetic variable to compile, lay out, and evaluate each step. Cutting the residual helpers directly reduces LTM compile time and per-step cost on large models.
• Maintainability. Several of these helper sources exist only because qualification lacks the per-variable declared dims at the wrap site -- threading that context through removes whole categories of fallback synthesis rather than papering over them.

Components affected

• src/simlin-engine/src/ltm_augment.rs -- wrap_index_non_matching_in_previous, qualify_element_index (the wrap path that lacks the subscripted variable's declared dims and so falls back to helper synthesis).
• src/simlin-engine/src/db/ltm/link_scores.rs -- the call sites that already thread per-variable declared dims into qualification (the pattern to extend to the wrap path).
• src/simlin-engine/src/builtins_visitor.rs -- index_is_static (:322) and the static-subscript check (:767) that gate direct LoadPrev/INIT compilation vs. helper synthesis.

Possible directions

• Thread per-variable declared dims into the wrap path so qualify_element_index can be position-matched everywhere -- the same approach already used at the link_scores.rs call sites. This removes the bulk of (1) and (3).
• Extend the builtins-visitor static check (index_is_static) to accept unambiguous bare element names, so they compile to a direct static-subscript PREVIOUS instead of a helper.
• Resolve module-output refs to static slot offsets (submodel-offset resolution -> direct LoadPrev) to eliminate (2).

How it was discovered

Follow-up work from the LTM compile-time/correctness work on 2026-05-31 (commits 7e6048e7, 82d20304, 51392894), profiling residual helper-aux synthesis on C-LEARN v77 with LTM discovery enabled after the model came under the slot limit.

Linkage

• Part of the LTM epic epic: Loops that Matter (LTM) -- tracking known issues and improvements #488.
• Relates to LTM: PREVIOUS-in-subscript arg not reduced to a bare var during helper rewriting -> LTM synthetic link/loop score silently drops to 0 #587 (subscripted-PREVIOUS LTM fragment drops to 0 -- now partially addressed) and ltm-finding: strongest-path discovery infeasible on large arrayed models (element-level blowup) #647 (element-level discovery blowup); this issue is the residual helper-synthesis volume / slot-ceiling-headroom concern, a distinct root cause from both.