ltm: module-internal PREVIOUS lag is uncounted state on every surface (lag-only sub-model loops unscored and unpinnable)

[bpowers]

Problem

Module-INTERNAL PREVIOUS() state is uncounted on every LTM surface. A loop

driver -> sub(out = PREVIOUS(in)) -> reader -> driver

is genuine feedback -- the sub-model implements a one-DT lag, the parent cycle compiles and simulates as a first-order discrete lag -- but it is entirely unscored and unpinnable:

• model_is_stateless's lagged leg (model_has_lagged_dt_deps, src/simlin-engine/src/db/ltm/mod.rs) is deliberately parent-level only: it checks previous_only dt deps on the model's own variables, skipping Module kinds.
• The module leg (modules_carry_state, same file, ~line 288) walks sub-models transitively but tests only sub_edges.stocks.is_empty() -- a sub-model whose only state is a PREVIOUS lag reads as stateless.
• Pin validation (src/simlin-engine/src/db/ltm/pinned.rs) mirrors the same blindness: it cannot see lagged edges inside a module's pathway, so a pin over the parent loop is rejected.

So on a parent whose only state is a lag-only sub-model, the stateless early return fires and LTM emits nothing (no scores, and pins are dropped without even an invalid-pin warning), while model_detected_loops -- which has no stateless gate -- still enumerates the loop. The detected-vs-scored gap is the same shape #748 describes for module-internal stocks, on the lagged-state leg.

This is a known, documented residual of the #749 fix (branch ltm-fix-batch-2, commit 69e84d8e): that fix made PARENT-level previous_only dt deps count as state in both model_is_stateless and pin validation, and deliberately left module-internal lagged state uncounted on BOTH surfaces so the two surfaces at least agree (both treat the shape as stateless). The rustdoc on model_has_lagged_dt_deps records the scope cut.

Why it matters

Genuine feedback is invisible to LTM for lag-only sub-models -- e.g. a user sub-model implementing a custom delay or smoothing structure via PREVIOUS instead of an INTEG. #749 established (per the LTM reference, sec. 7) that PREVIOUS is state the enumerator legitimately scores; the module boundary should not change that verdict. Severity: low-medium (uncommon model shape today; silent no-output rather than wrong numbers).

Components affected

• src/simlin-engine/src/db/ltm/mod.rs -- modules_carry_state (~line 288) and/or model_has_lagged_dt_deps (the stateless gate's two legs)
• src/simlin-engine/src/db/ltm/pinned.rs -- pin validation's view of module-internal lagged edges

Constraint / fix direction

Both surfaces must move together (the #749 invariant): give modules_carry_state a lagged leg (a sub-model variable with a previous_only dt dep counts as state, transitively, mirroring its stock walk) AND give pin validation module-pathway lagged-edge visibility, so a scored loop is also pinnable. Fixing only the gate would recreate the exact enumerable-but-unpinnable disagreement #749 closed.

Cross-references

• ltm: stockless PREVIOUS-lagged cycle is scored by the enumerator but rejected by pin validation (false 'no stock' rationale) #749 -- the parent-level fix whose documented scope cut this issue tracks (commit 69e84d8e on ltm-fix-batch-2)
• ltm: model_ltm_variables' stock-free early return is module-blind -- a root whose only state is module-internal silently emits no LTM variables #748 -- the module-blind stock-free early return (stock leg of the same gate; same silent-no-output family)
• LTM: module stock enrichment misses stocks reachable only through internal feedback cycles #282 -- enrichment misses cycle-only module-internal stocks (another module-visibility gap, different code site)
• ltm: loops through SMOOTH/DELAY/TREND builtin calls are unpinnable (synthetic module node unnameable and unbridgeable by order_variable_cycle) #747 -- builtin-call loops unpinnable (module-boundary pin gap for synthetic instances)
• LTM: improve/generalize module support #274 -- module-support umbrella; part of epic epic: Loops that Matter (LTM) -- tracking known issues and improvements #488

Discovery context

Identified during the #749 fix on branch ltm-fix-batch-2 (commit 69e84d8e), 2026-06-09: the commit message and model_has_lagged_dt_deps rustdoc record the deliberate scope cut; this issue makes the residual independently trackable.