v1.7.5 sequential-learning adapter contract + multi-seed eval (eval STOP)

[kitfunso]

Summary

Sequential-learning benchmark infrastructure release. Closes the v0.39 B3 follow-up that gated public exercising of the dlPFC goal-stack mechanism. Eval ran but stopped per pre-registered sanity gate due to floor effect; hypothesis remains open.

• Task 1 — Adapter contract pushGoal / completeGoal hooks on interface.mjs. hippo.mjs implements both via existing hippo goal push|complete CLI with HIPPO_HOME / XDG_DATA_HOME isolation. Tag-fix on memory store ([task.trapCategory, ...category.tags, 'error']) so the boost can match — without it the eval would have RETRACTED a working mechanism.
• Task 2 — Multi-seed eval harness with category-to-slot variance (seeded mulberry32, hash-derived sub-seeds per shape group). Permutation CI for paired Δ (exact sign-flip, 10k resamples). --seed, --n-seeds, --eval-strict flags.
• Task 3 — Pre-registered protocol + claim inventory committed BEFORE eval. 4 conditions × 20 seeds × eval-strict. Sanity gate fired (C2 late = 0% << headline 14%); STOP applied per prereg. Floor effect — both C2 and C3 saturate at 0% late-phase. ΔLate = 0pp. Mechanism shipped, hypothesis open.

Outside-voice review payoff

Plan v1 → 13 P0/P1 findings from senior-code-reviewer + Codex CLI. v2 plan addressed all. Biggest catches: tag mismatch (boost would have matched zero memories), within-thirds shuffle was statistical theatre, paired t-test fragile for bounded discrete rates → permutation CI, HIPPO_HOME / XDG_DATA_HOME isolation, hard-fail in eval-strict mode.

Test plan

• npx vitest run → 1445 passed (+27 new), 0 failures
• npx tsc --noEmit → 0 errors
• npm run build → clean
• Pre-eval sanity: tag-fix verified to fire boost (Task 1 integration test)
• Eval ran cleanly: zero hook failures across all 4 conditions × 20 seeds in eval-strict mode

Eval result

STOP per pre-registered sanity gate. C2 (hippo-base) late-phase = 0.0% across all 20 seeds, outside the pre-registered band [4%, 24%]. Both C2 and C3 saturate at 0% late — floor effect prevents H1/H0 discrimination. The −10pp hypothesis remains untested on a discriminating workload. Documented in docs/evals/2026-05-07-v1.7.5-goal-stack-eval-result.md with full investigation. Future eval needs harder benchmark variant (smaller --budget, adversarial categories, or restricted late-phase window).

Plan: docs/plans/2026-05-07-v1.7.5-sequential-learning-adapter.md.

Closes the v0.39 B3 follow-up "sequential-learning adapter contract" in TODOS.md. The "−10pp hypothesis" line stays open as v1.7.6+ followup.