research: fold open PR branches into the honest-scholar hypothesis workflow

[davorrunje]

Context

mononet was onboarded onto the honest-scholar workflow in #136 (research-init adopt), which established docs/research/ with four papers and retroactive hypotheses. Several in-flight feature branches / draft PRs predate the onboarding and are not yet expressed as honest-scholar hypotheses, so their work does not enter through the hypothesis-testing front door and their results are not wired as findings.md evidence.

Design context: docs/superpowers/specs/2026-07-21-honest-scholar-integration-design.md (§6, deferred) and .honest-scholar/config.yml.

Goal

Map each open feature PR/branch to a hypothesis under the correct paper in docs/research/, so in-flight engineering resumes through the honest-scholar hypothesis-testing flow and its committed results attach as evidence.

Where

• Papers: docs/research/monotone-constructions/, docs/research/structure-preserving-pinns/.
• Existing retroactive hypotheses to attach to (some already stubbed, unsigned):
  • docs/research/monotone-constructions/hypotheses/2026-07-21-alternate-construction-base/ (open draft PR)
  • docs/research/monotone-constructions/hypotheses/2026-07-21-deep-monotonic-residual-accuracy/ → PR feat(benchmarks): deep monotonic residual accuracy flavor (Stage 2 plumbing) #114 (feat/deep-residual-accuracy)
  • docs/research/monotone-constructions/hypotheses/2026-07-21-hp-search-sensitivity/ (not done)
  • docs/research/monotone-constructions/hypotheses/2026-07-21-flavor-convex-fraction-ablation/ (not done)
• Open PRs/branches to triage (as of 2026-07-21): Handoff: monotone-constructions (mixed/alternate/split) — resume point #135 (docs/monotone-constructions-handoff), Applications: Structure-Preserving PINNs (Paper 1) — framework + inverse/traffic headline #116 (spec/applications-structure-preserving-pinns), bench(results): large-dataset screen — monotone depth is neutral across all 5 datasets #115 (feat/screen-results-clean), feat(benchmarks): deep monotonic residual accuracy flavor (Stage 2 plumbing) #114 (feat/deep-residual-accuracy), bench: larger synthetic n_train + L=4 Stage-2 recast #113 (chore/synth-larger-for-large-batch-band); branches feat/alternate-construction, feat/alternate-base-result, feat/phase4-flavor-ablation, feat/stage2-unified-depth-benchmark, feat/monotone-depth-probe.

Proposed approach

For each open PR/branch: identify the hypothesis it advances (create the folder from templates if none), set findings.md evidence: to the result JSONs the branch commits, and once the analysis completes, obtain the author's named sign-off on the verdict. PR #115 (large-dataset screen: "monotone depth is neutral across all 5 datasets") is the natural evidence for deep-monotonic-residual-accuracy.

Acceptance criteria

• Every currently-open feature PR/branch is either mapped to a hypothesis folder or explicitly closed/won't-do.
• The four unsigned hypotheses above have their evidence: wired to committed run-refs (or remain honestly pending with a blocker noting why).
• docs/research/dashboard.md regenerated to reflect the new mappings.

References

• chore: adopt honest-scholar as the research-workflow driver (install + research-init adopt) #136, Design spec: benchmark experiment orchestration #127, feat(benchmarks): deep monotonic residual accuracy flavor (Stage 2 plumbing) #114, bench(results): large-dataset screen — monotone depth is neutral across all 5 datasets #115, Applications: Structure-Preserving PINNs (Paper 1) — framework + inverse/traffic headline #116, Handoff: monotone-constructions (mixed/alternate/split) — resume point #135
• docs/superpowers/specs/2026-07-21-honest-scholar-integration-design.md