ltm-augment: TIME-bearing reducer bodies break the PREVIOUS(agg) anchor in body-aware partials (nonlinear arm attributes full time-drift to every row)

[bpowers]

Summary

The body-aware per-row source[e] -> $⁚ltm⁚agg⁚{n} partials break the PREVIOUS(agg) anchor identity whenever the reducer body references TIME, a time builtin (PULSE/STEP/RAMP), or a nested PREVIOUS(x). freeze_pinned_body (src/simlin-engine/src/ltm_augment.rs ~3457) freezes only model dependencies, and reducer_body_ctx_parts (src/simlin-engine/src/db/ltm/link_scores.rs ~975) deliberately excludes TIME from the dep set -- so in a "fully frozen" term, TIME (and the time builtins, and nested PREVIOUS(x)) stays LIVE. A row whose every model dep is frozen therefore does NOT reproduce the previous-step value: R(all-frozen terms) != PREVIOUS(agg).

On the NONLINEAR arm (MIN/MAX/STDDEV), the partial's anchor subtraction R(live row, frozen rows) - PREVIOUS(agg) then attributes the full time-drift to EVERY row -- including provably non-influential ones -- destroying the frozen-argmin-scores-0 property.

Repro (demonstrated at eab752c8 on ltm-fix-batch-2)

grow = 1 + MIN(pop[*] * TIME)

The south row is never the argmin (its true ceteris-paribus partial is 0 at every step), yet it scores 0.9999, 0.9920, 0.9764, ... -- the drift pop_north_prev * delta-TIME duplicated onto every row.

Scope by arm

• NONLINEAR (MIN/MAX/STDDEV): the anchor subtraction duplicates the drift onto every row -- demonstrated wrong-value class above.
• LINEAR (SUM/MEAN, the ltm: source->agg link score wrong-signed for hoisted multi-source reducer with sign-flipping coefficient (SUM linear shortcut ignores the body's coefficient) #744 fix in 5a533bf6): the same deps-only-freeze convention only scales the live row's own delta (the per-row partial is a difference of two terms that share the frozen co-rows), so the effect is qualitatively benign there; the nonlinear arm's anchor identity is what breaks.

Why it matters

Low-medium severity: silently wrong per-row link scores (and the loop scores through those rows inherit the drift) for time-bearing reducer bodies -- an uncommon but perfectly legal equation shape. NOT a regression: the pre-#762 bare substitution emitted worse garbage for the same body, and the delta-ratio fallback would also score ~1 here. But it is a demonstrated wrong-value class that was invisible in the rustdoc (a doc caveat citing this issue is being added on the branch).

Component

• src/simlin-engine/src/ltm_augment.rs (freeze_pinned_body, the MIN/MAX/STDDEV builders in generate_nonlinear_partial)
• src/simlin-engine/src/db/ltm/link_scores.rs (reducer_body_ctx_parts's TIME exclusion)

Fix directions

1. Bail to delta-ratio when the pinned body references TIME, a time builtin, or PREVIOUS -- detectable during pinning. Honest relabeling (delta-ratio is the documented conservative fallback) rather than a better value.
2. Freeze time-like terms too by anchoring both evaluations of the guard form at the SAME time sample, so the anchor identity R(all-frozen) = PREVIOUS(agg) holds again. A deeper change to the guard form, but recovers the frozen-argmin-scores-0 property.

Cross-references (related, NOT duplicates)

• ltm-augment: generate_nonlinear_partial is body-blind for non-bare MIN/MAX/STDDEV reducer bodies (garbage per-row link scores) #762 -- the body-aware nonlinear-partial work this limitation was found in (non-bare bodies generally); this issue is the time-bearing residue of that fix.
• ltm: source->agg link score wrong-signed for hoisted multi-source reducer with sign-flipping coefficient (SUM linear shortcut ignores the body's coefficient) #744 -- the linear sibling whose pin_body_to_row/freeze_pinned_body machinery (commit 5a533bf6) carries the deps-only-freeze convention.
• ltm-augment: instantaneous and flow-to-stock link scores use inconsistent time windows within a loop #544 -- intra-loop time-window inconsistency; same "which time sample does the guard form see" family, different code site.
• Epic epic: Loops that Matter (LTM) -- tracking known issues and improvements #488 (Augmentation group).

Discovery context

Demonstrated during adversarial review of the GH #762 work on branch ltm-fix-batch-2 (commit eab752c8).