LTM: PREVIOUS-in-subscript arg not reduced to a bare var during helper rewriting -> LTM synthetic link/loop score silently drops to 0

[bpowers]

Summary

A PREVIOUS(...) call whose argument sits inside a Subscript index (e.g. PREVIOUS(self[COP, tNext])) is not reduced to a bare Expr::Var by the PREVIOUS helper-rewrite pass. When such an expression appears in an LTM synthetic link-score / loop-score fragment (generated only when LTM discovery is enabled), codegen rejects it and the fragment is silently dropped, so that LTM link/loop score reads a constant 0 -- degrading LTM loop-discovery accuracy for that edge.

This is not a simulation-correctness bug: core simulation is unaffected. It is an LTM-accuracy gap on the loop-analysis path. It is also not a blocker for the element-cycle-resolution work -- C-LEARN compiles, runs, and the LTM-discovery-compiles checks all pass; this surfaces (loudly, via a Warning) once that work removed the masking failures.

Failure chain

1. The PREVIOUS helper-rewrite (src/simlin-engine/src/builtins_visitor.rs, the scalar-helper synthesis for non-bare PREVIOUS args -- walk() / make_temp_arg) does not reduce a PREVIOUS argument that is an Expr::Subscript index to a bare per-element scalar Expr::Var. The argument survives helper rewriting still shaped as a Subscript.
2. When this expression is part of an LTM synthetic fragment, codegen reaches the Expr::Subscript(off, indices, ...) -> SubscriptIndex::Single(expr) arm at src/simlin-engine/src/compiler/codegen.rs (~:489-:521, the self.walk_expr(expr)? at ~:493-:502). The inner walk returns Err(NotSimulatable, "PREVIOUS requires a variable reference after helper rewriting"), which -- as of commit 59f01126 -- is propagated via ? rather than panicking. (That ? propagation is exactly the Incremental compiler panics on C-LEARN model (catch_unwind workaround) #363 / AC7.5 fix: the in-tree comment at codegen.rs:494-501 documents "a PREVIOUS whose arg survived helper rewriting as a non-variable expression (NotSimulatable) ... must flow back to the caller (db_ltm.rs gracefully drops the un-compilable LTM synthetic fragment), never escalate to a panic (Incremental compiler panics on C-LEARN model (catch_unwind workaround) #363).")
3. That Err flows back to compile_phase in src/simlin-engine/src/db_ltm.rs, where the Err(_) => None arm at db_ltm.rs:2083 (and the sibling Err(_) => None build_var fallbacks at ~:2093/:2099) gracefully drops the un-compilable LTM synthetic fragment.
4. With no bytecode emitted for that synthetic variable, it keeps its layout slot and reads a constant 0 for the entire simulation. The loop score for any loop through that edge is the product of its link scores, so a zero link score collapses that loop's score.

The drop is surfaced, not silent: the existing model_ltm_fragment_diagnostics pass emits a Warning for every LTM synthetic variable whose fragment fails to compile (subject to the #466 caveat that the diagnostic-collection FFI paths leave ltm_enabled false).

Why it matters

• LTM accuracy. A subscripted-self idiom such as PREVIOUS(self[COP, tNext], init) (the C-LEARN SAMPLE IF TRUE-style construct) is legal and appears in real models. When an edge using it is in a detected loop and gets scored under LTM, its link/loop score silently reads 0, so the loop looks inactive when it is not -- the same hazard class as ltm: synthetic-fragment compilation uses an empty tables map, so link scores into graphical-function targets stub to 0 #546 (GF-table targets stub to 0) and ltm: composite-reference link score into a stdlib-macro module fails to compile in exhaustive mode, so loops through SMOOTH/DELAY stub to 0 #548 (composite-reference into a stdlib macro stubs to 0).
• Bounded blast radius. Core simulation is unaffected (this is the LTM-augmentation path only), and the failure is no longer a panic (Incremental compiler panics on C-LEARN model (catch_unwind workaround) #363/AC7.5) -- it is a clean drop with a Warning. So this is a correctness/accuracy gap in LTM scoring, not a crash or a sim-result regression.

Repro context

Surfaced and made loud during element-cycle-resolution Phase 6 Task 11 (#363 / AC7.5), commit 59f01126, on test/xmutil_test_models/C-LEARN v77 for Vensim.mdl with LTM discovery enabled.

Components affected

• src/simlin-engine/src/builtins_visitor.rs -- the PREVIOUS helper-rewrite (walk() / make_temp_arg): a subscripted/expression PREVIOUS argument is not reduced to a bare per-element scalar Expr::Var.
• src/simlin-engine/src/compiler/codegen.rs -- the Expr::Subscript -> SubscriptIndex::Single arm (~:489-:521) where the inner walk_expr(expr)? returns the NotSimulatable "PREVIOUS requires a variable reference after helper rewriting" error.
• src/simlin-engine/src/db_ltm.rs -- db_ltm.rs:2083 (Err(_) => None, plus the sibling build_var Err(_) => None arms), the graceful-drop site that makes the LTM synthetic fragment read constant 0.

Possible fix direction

Extend the PREVIOUS helper-rewrite in builtins_visitor.rs to reduce a subscripted (or otherwise non-bare-expression) PREVIOUS argument to a bare Expr::Var -- i.e. synthesize the per-element scalar helper aux for the subscripted slice -- so the LTM synthetic fragment compiles instead of being dropped. This is the same structural direction proposed for the related PREVIOUS-arg cases (generalize the helper synthesis so PREVIOUS(expr) desugars for any expr, not just a bare Var). The fix should be paired with a regression test asserting the LTM link/loop score for an edge through a subscripted-PREVIOUS reference is non-zero (so the dropped-fragment-reads-0 behavior cannot silently return).

Linkage

• Part of the LTM epic epic: Loops that Matter (LTM) -- tracking known issues and improvements #488.
• Lineage: Incremental compiler panics on C-LEARN model (catch_unwind workaround) #363 / AC7.5 (the element-cycle-resolution change that made the ?-propagation path -- and therefore this graceful drop -- the live behavior; commit 59f01126).
• Same hazard class ("LTM fragment compiler can't handle a particular shape -> synthetic var stubs to 0 -> degraded score masquerades as inactive, surfaced only by the model_ltm_fragment_diagnostics Warning") as ltm: synthetic-fragment compilation uses an empty tables map, so link scores into graphical-function targets stub to 0 #546 (GF-table targets) and ltm: composite-reference link score into a stdlib-macro module fails to compile in exhaustive mode, so loops through SMOOTH/DELAY stub to 0 #548 (composite-reference into a stdlib macro) -- but a distinct root cause and codegen site.
• Distinct from 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: 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 is the partially-iterated arrayed subscript (pop[Region, young]) classified DynamicIndex, where the A2A partial generator wraps a SUM(...) (an Expr::App) in PREVIOUS, failing at the BuiltinFn::Previous arm (codegen.rs ~:701/:712). This issue's failing argument is an Expr::Subscript (not an App) and the error fires at the Expr::Subscript -> SubscriptIndex::Single arm (codegen.rs ~:493), via the subscripted-PREVIOUS helper-rewrite path -- not the partial generator's SUM-wrap.
• Distinct from engine: bare arrayed name inside nested PREVIOUS() in an apply-to-all equation fails to compile #541: engine: bare arrayed name inside nested PREVIOUS() in an apply-to-all equation fails to compile #541 is a bare arrayed name inside a nested PREVIOUS(PREVIOUS(pop)), a non-LTM whole-project hard compile failure rooted in make_temp_arg synthesizing a scalar helper for a bare arrayed name (no subscript to substitute). This issue's argument is a subscripted reference, and the failure is the graceful LTM-fragment drop (score -> 0), not a hard whole-project failure.
• Distinct from C-LEARN: post-cycle-gate assemble_simulation fragment-compile failure on synthetic temp-arg helpers #580: C-LEARN: post-cycle-gate assemble_simulation fragment-compile failure on synthetic temp-arg helpers #580 is the C-LEARN post-cycle-gate assemble_simulation NotSimulatable failure on the ~150 real/synthetic temp-arg helper model variables (the core-compile path returning Err). This issue is specifically the LTM synthetic link/loop-score fragment being gracefully dropped at db_ltm.rs:2083 (the LTM-discovery path), which leaves the model simulating fine with a degraded loop score -- not a NotSimulatable core-compile failure.

[bpowers]

Fixed by 865854c ("engine: compile static-subscript PREVIOUS/INIT directly, qualify LTM element refs" — its message names this bug) and f1ac6f6. PREVIOUS/INIT with statically-resolvable subscript indices (including A2A-iterated dimension elements, the SAMPLE IF TRUE self-reference) now compile to a direct LoadPrev/LoadInitial, and the ceteris-paribus wrapper no longer PREVIOUS-wraps an element index; the dynamic-index case keeps a helper aux that itself compiles.

Verified empirically on main (521fc37): a minimal 2D-arrayed SAMPLE IF TRUE-in-loop model compiles with zero fragment-compile-failure warnings and a non-zero link/loop score through the construct; test_previous_a2a_iterated_dimension_no_helpers, test_ltm_bare_element_subscripts_no_helpers, and test_previous_dynamic_subscript_keeps_helper pass; and on C-LEARN no remaining fragment failure involves the SAMPLE IF TRUE / self-subscript shape (the remaining C-LEARN fragment failures belong to #528 and module-composite cases, tracked separately).