Add asymmetric_translog: translog with sign-split own squares around an estimated knot

[hmgaudecker]

Stacked on #103 (missing-observed-factors). Replaces #101: stable_translog is dropped; this PR also carries #101's CI-sharding commit.

Implements docs/superpowers/specs/2026-10-16-asymmetric-translog-design.md.

What

New built-in transition function

f(x) = c + Σ β_j x_j + Σ γ⁺_j max(x_j − κ_j, 0)² + Σ γ⁻_j min(x_j − κ_j, 0)² + Σ_{a<b} δ_ab x_a x_b

• One curvature coefficient per side of a per-factor knot κ_j; C¹ in x and κ (no where, no clipping).
• Knots are parameters, estimated by default; fix them through the usual machinery (e.g. to profile over a grid).
• Nests translog: γ⁺ = γ⁻ gives a translog with shifted linear/constant terms; with κ = 0 it is exactly translog.
• Parameter layout: linear, f ** 2 (above), f ** 2 (below), f knot, interactions, constant. Identity constraints: own linear 1, everything else (knots included) 0.

Motivation (spec §1): in the health-cognition application the measures revert strongly at one end of the range and are near-linear at the other. A single square coefficient then forces an own slope above one at the near-linear end; a symmetric clip of the states cannot separate the two ends either.

Registration

_AF_COMPATIBLE_TRANSITIONS, _LEAKY_BUILTIN_PRODUCTION (AF), builtin_names (AMN resolver), the reference-guide lists.

Tests

• tests/test_asymmetric_translog.py: translog nesting (κ = 0 and κ ≠ 0, 1e-12), one-sided derivatives, C¹ at the knot in x and κ, names/identity constraints, JIT/vmap/50-step scan vs eager, and a fixed-point witness for the application pattern.
• tests/test_asymmetric_translog_dispatch.py: public model, AMN resolver, AF allow-lists/validation, with a value that distinguishes the two sides of the knot.
• tests/test_asymmetric_translog_estimation.py: a registered same-law reference reproduces CHS loglikeobs (1e-10) and the gradient (rtol 1e-8); equal curvature with zero knots reproduces translog; unequal curvature and swapped sides change the result; endogenous carry-forward substeps accept the function.

Full suite (pytest tests -m 'not long_running' -n 2 on 831db22): 857 passed.

https://claude.ai/code/session_017psXwkMuDeVmTu5WgoM26k

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.51%. Comparing base (9cd73db) to head (1be29c7).

Additional details and impacted files

@@                 Coverage Diff                 @@
##           stable-translog     #102      +/-   ##
===================================================
+ Coverage            95.42%   95.51%   +0.08%     
===================================================
  Files                  148      151       +3     
  Lines                15178    15459     +281     
===================================================
+ Hits                 14484    14765     +281     
  Misses                 694      694              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB