Handoff: monotone-constructions (mixed/alternate/split) — resume point

[davorrunje]

Draft handoff / bookmark. Not for merge — a resume point for the alternate-construction program so a later session can pick it up cold. Full note: docs/superpowers/plans/2026-07-21-monotone-constructions-continuation.md.

TL;DR

The alternate construction + composition-aware prev= init are built, merged, CI-green (3 backends). The one experiment that actually tests why alternate exists — the deep plain verdict — has not been run. Everything measured so far is either shallow (parity) or residual (a wash). Finish the deep-plain grid run before concluding.

Merged (on main)

• Monotone constructions: mixed/alternate/split rename, composition-aware init, and flavor ablation (design + phase-1 plan) #108 phase 1 — rename mode absolute→mixed, switch→split (hard break, no behavior change).
• feat: alternate monotonic construction + composition-aware prev= init (Phase 2) #109 phase 2 — alternate construction + composition-aware init via prev= (torch/jax/keras). API live: MonoLinear/MonoDense(mode="alternate", prev=<prev|None>); layer-resolved to the |W| kernel; convex_fraction reserved for mixed; MonoResidual rejects alternate.
• feat(bench): flavor-ablation benchmark (Phase 4) — design + plan for GPU execution #110 phase 4 — flavor-ablation harness + plan (benchmarks/flavor_ablation.py, plain depths {4,8,16}, divergence-rate).
• feat(bench): tuned shallow base result — 4-flavor bake-off + convex_fraction ablation #111 tuned shallow (≤4-layer) 4-flavor bake-off + convex_fraction ablation. feat(bench): HP-search sensitivity curves (spec + plan + implementation) #117 HP-sensitivity curves.

Evidence so far

• Shallow bake-off (feat(bench): tuned shallow base result — 4-flavor bake-off + convex_fraction ablation #111): at ≤4 tuned layers, alternate does not decisively beat mixed/split (co-leads compas, parity auto/heart, loses blog/loan). Biggest win was fixing convex_fraction=0.5 in mixed (3/5).
• Deep residual screen (bench(results): large-dataset screen — monotone depth is neutral across all 5 datasets #115, draft): "monotone depth is neutral" — the H-residual wash regime.
• CPU prior (spec §6): at depth-16 plain, mixed/split diverge while alternate stays stable → the untested claim.

The outstanding experiment — deep plain (H-plain)

Not covered by #111 (shallow) or #114/#115 (residual). The grid runner from #110 is built for it (plain, depths {4,8,16}, divergence-rate) but has not run — no benchmarks/results/flavor-ablation/ on main. Run commands + how to read the verdict are in the handoff note. Verify the runner CLI matches the #110 plan before running.

Other open threads

• Fix convex_fraction=0.5 default (feat(bench): tuned shallow base result — 4-flavor bake-off + convex_fraction ablation #111 follow-up).
• Phase 3 — concepts-docs distillation (docs/concepts/, CPU, not started).
• Notebook release-gate — tools/execute-benchmarks.sh (GPU) to refresh .ipynb outputs before 0.0.0a1.
• Deferred cleanups — absolute_init_params→mixed_init_params, equivalence filename slugs.

Pointers

Spec 2026-07-13-monotone-constructions-init-and-ablation-design.md; benchmark spec 2026-07-14-flavor-ablation-benchmark-design.md; plans 2026-07-14-{mode-rename-migration,phase2-alternate-construction,phase4-flavor-ablation}.md; results benchmarks/results/alternate-base/ + docs/benchmarks/alternate-base-result.md; related in-flight #113/#114/#115; auto-memory alternate-construction-init.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (50c3232) to head (a92ee33).

Additional details and impacted files

@@            Coverage Diff            @@
##              main      #135   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           18        18           
  Lines          867       867           
  Branches       148       148           
=========================================
  Hits           867       867           

🚀 New features to boost your workflow:

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