epic: Loops that Matter (LTM) -- tracking known issues and improvements

[bpowers]

Purpose

Single source of truth for known open work in the Loops that Matter (LTM) subsystem. Use this issue to find a bite-sized piece to work on without re-investigating the full subsystem state.

For a one-paragraph description of LTM, see src/simlin-engine/CLAUDE.md and docs/architecture.md. The reference paper and design notes live under docs/design/ and docs/design-plans/.

How to use this tracking issue

• Looking for something to pick up? Browse the grouped checklists below and pick an unchecked sub-issue. Each sub-issue has the ltm label and stands alone with file:line references and a suggested fix.
• Filed something new? Add it to the appropriate group below as a tasklist item (- [ ] #N -- short title). GitHub auto-renders the issue's open/closed state.
• Closed something? Tick the box here too -- the box is the human-readable progress signal even though the underlying issue state is the source of truth.
• Filtering on the command line: gh issue list --label ltm --state open returns the full open set.

Recently resolved (2026-04-29 triage)

• ltm: extend composite module scoring beyond implicit stdlib modules #306 -- composite module scoring extended to user-defined dynamic modules (commit 01970ba2)
• ltm: black-box module link scoring collapses magnitude to 1 #307 -- black-box module link score now uses true delta-ratio (commit c16d9a90)
• engine: CrossElement over-approximation for fixed-index subscript refs in LTM causal graph #448 -- FixedIndex N x N over-approximation replaced by per-reference RefShape emitter (commit 1272dd04)
• engine: gate wrld3_element_level_enumeration_is_uncapped behind cfg(not(debug_assertions)) #465 -- wrld3_element_level_enumeration_is_uncapped gated with #[ignore] (commit fee04c61)
• LTM layout: add integration/unit coverage for detected.truncated fallback branch #467 -- detected.truncated fallback branch removed; auto-flip backstop covers the dense-graph case (commit 88ca6498)

Open work

Grouped by code area. Each row links to a GitHub issue with full context.

Core algorithm -- src/simlin-engine/src/ltm.rs, src/simlin-engine/src/db_analysis.rs

• ltm: dedup by node set merges distinct directed loops #308 -- dedup by sorted node set merges distinct directed cycles (key on canonical edge-sequence rotation instead) [medium]
• ltm: JohnsonState per-SCC allocation sized to whole-graph node count #460 -- JohnsonState per-SCC allocation sized to whole-graph node count instead of scc.len() [low; only matters for sparse-SCC graphs]
• ltm: fixed max depth 20 truncates module pathway enumeration #312 -- enumerate_module_pathways truncates at hard-coded depth 20 [low; affects deeper user modules]
• ltm: polarity analysis returns Unknown for all array reducers #480 -- polarity analysis returns Unknown for all array reducers (SUM, MEAN, STDDEV, RANK, MIN/MAX over array) [medium]
• ltm: enumerate variable-level loops first, expand only cross-element subgraph #482 -- enumerate variable-level loops first, then expand only the cross-element subgraph [medium; pure-A2A models trip auto-flip prematurely]
• ltm: surface polarity confidence ratio for Rux/Bux classification #485 -- surface polarity confidence ratio for Rux/Bux classification [low; capability gap]

Discovery / post-sim -- src/simlin-engine/src/ltm_finding.rs, src/simlin-engine/src/ltm_post.rs

• ltm-finding: discovery runs at save_step instead of dt #309 -- discovery iterates at save_step granularity instead of dt; sub-save_step transients are missed [medium]
• ltm-finding: truncate-before-filter can drop partition-dominant loops #310 -- rank_and_filter truncates to MAX_LOOPS=200 before partition-aware filtering; small-partition dominant loops can be lost [low; comment in code already acknowledges]
• ltm-finding: remove hardcoded 'main' model assumption #313 -- find_main_model is hardcoded to canonical \"main\"; discovery cannot run on other top-level models [low; usability]
• LTM post-sim: preserve emission order for rel_loop_score denominator summation #468 -- LtmVariablesResult.loop_partitions is HashMap; post-sim helpers .sort() IDs lexicographically (b1, b10, b11, ..., b2) instead of preserving emission order [low; bit-level drift only]
• ltm-finding: SearchGraph clones Ident<Canonical> in per-timestep DFS hot path #481 -- SearchGraph clones Ident<Canonical> in per-timestep DFS hot path; switch to NodeId indexing for 5-10x speedup [medium; perf]
• ltm: implement LOOPSCORE and PATHSCORE builtins for user-pinned loops #484 -- implement LOOPSCORE / PATHSCORE builtins so users can pin specific loops/paths [low; capability gap]
• ltm: rel_loop_score normalisation cross-pollutes independent A2A loops #487 -- rel_loop_score normalisation cross-pollutes independent A2A loops because partition_for_loop returns None [medium; silent correctness for multi-A2A-loop models]

Augmentation -- src/simlin-engine/src/ltm_augment.rs

• ltm-augment: parse fallback can silently break ceteris-paribus logic #311 -- build_partial_equation_shaped parse-failure fallback returns lowercased text without PREVIOUS() wrapping, silently breaking ceteris-paribus [medium; silent correctness on edge cases]
• ltm: LTM_BENCH_TRACE should be cfg(any(test, feature = "ltm_bench")) instead of runtime env check #464 -- LTM_BENCH_TRACE runtime env-var check; production builds carry tracing eprintlns and read_rss_mib [low; build hygiene]
• ltm-augment: STDDEV/RANK partials silently fall back to delta-ratio #483 -- STDDEV and RANK ceteris-paribus partials silently fall back to delta-ratio (MIN/MAX are now correct) [medium; silent correctness]
• ltm: emit compile-time error when ltm_enabled is true and integration is non-Euler #486 -- emit compile-time error when ltm_enabled is true and integration method is non-Euler [medium; silent correctness when both enabled]

Module / array umbrellas + FFI

• LTM needs to support arrays #273 -- LTM array support (umbrella). Substantial progress; remaining sub-gaps are tracked individually as ltm: polarity analysis returns Unknown for all array reducers #480, ltm: enumerate variable-level loops first, expand only cross-element subgraph #482, ltm-augment: STDDEV/RANK partials silently fall back to delta-ratio #483, ltm: rel_loop_score normalisation cross-pollutes independent A2A loops #487 [umbrella]
• LTM: improve/generalize module support #274 -- LTM module support (umbrella). Composite scoring now generalises to user-defined dynamic modules (ltm: extend composite module scoring beyond implicit stdlib modules #306 closed); pathway depth-20 cap (ltm: fixed max depth 20 truncates module pathway enumeration #312) and cycle-blind simple-path DFS (LTM: module stock enrichment misses stocks reachable only through internal feedback cycles #282) remain [umbrella]
• LTM: module stock enrichment misses stocks reachable only through internal feedback cycles #282 -- enrich_with_module_stocks simple-path DFS misses stocks reachable only through internal feedback cycles [low; rare topology]
• LTM auto-flip warning unreachable through simlin_project_get_errors #466 -- LTM auto-flip warning unreachable through simlin_project_get_errors because diagnostic collection compiles with ltm_enabled=false [medium; UX]

Cross-reference: GH issue ↔ docs/tech-debt.md item

GH issue	Tech-debt item
#308	#22 (same problem)
#310	#28 (same problem)
#480	#21
#481	#24
#482	#25
#483	#27
#484	#29
#485	#30
#486	#31
#487	#35

Items #20, #23, #26, #32, #34 in docs/tech-debt.md are already marked RESOLVED (kept as historical pointers to the fixing commits).

Conventions

• New LTM-related bugs/improvements: file as a normal issue with the ltm label and add a tasklist entry above.
• Use component-prefixed titles (ltm:, ltm-finding:, ltm-augment:, ltm-post:) so the issue list reads cleanly when grouped.
• When closing a sub-issue, tick the box here in the same PR (or a fast-follow) so the human view stays in sync.