Skip to content

Roadmap — quarantined experiments

Two uncalled Zenith-session prototypes were removed from main (refactor/kymora) and preserved on experiment branches. They are not deleted; they are not wired. Revival is gated on measurement per docs/internal/arch.md — "decide, then build".

PRD legacy (migrated 2026-10-04 from PRD.md, file removed)

Still-true post-1.0 ideas from the v1.0 PRD (§15), not otherwise tracked:

  • Polars output option alongside pandas (extract_features_df equivalent). Portability row in docs/internal/arch.md §4.5 already requires Polars inputs to work without copy where layout allows; a dedicated Polars output frame remains open.
  • R/Julia bindings — exploratory only, gated on demonstrated community demand (consistent with docs/internal/arch.md Z-GPU demand-gating posture).
  • Launch-plan content (PRD §§6, 14, 17–18: install/star targets, GTM phases, open questions) was dropped as dated, not migrated — §17 implementation items are all shipped (error wrapping, __version__, .pyi stubs, extract_features_df, property tests, NaN-policy tests, wheel matrix, CHANGELOG, CoC, templates). Product definition and non-goals already live in docs/internal/arch.md §1.

experiment/spin-pool — persistent spin-then-park pool (NEEDS-OWNER #7)

  • Contents: src/pool.rs (222 lines) as of f7282a0; was declared as pub mod pool in src/lib.rs with zero callers. TSXTRACT_POOL (pre-rename name; now KYMORA_POOL) is read by nothing; tune.py's pool dimension benchmarks variants of a backend that does not exist (recorded as provisional in docs/api.md).
  • Revival gate — docs/internal/arch.md Z6 (§10): re-baseline shows parallel efficiency η < 0.85 at (1k×500), or rayon wake/steal dominates the flamegraph. On revival, Z6 also requires: KYMORA_POOL=rayon fallback, panics converted to KymoraError, loom/stress tests + env switch (docs/internal/arch.md §12 pool row), and the §9.4 stop rule (stop after two consecutive attempts each gaining <3% end-to-end).
  • Revive with: git checkout experiment/spin-pool -- src/pool.rs, then wire + measure. Until then, tune() pool recommendations stay void.

experiment/soa-4x — lane-batched SoA kernels (NEEDS-OWNER #8)

  • Contents as of f7282a0:
  • SoA quartet aos_to_soa_4x / pass1_soa_4x / pass2_soa_4x / autocorr_soa_4x (src/kernels/reduce.rs:821–1153), zero callers;
  • run_core33_f32_out32 (src/pipeline.rs:252–260), zero callers.
  • None of the above is parity-tested (tests/ has zero references) and no benchmark demonstrates the ≥10% end-to-end gain that would have kept it on main behind a feature flag — so it moved here instead.
  • Revival gate — docs/internal/arch.md Z2 (§10): re-baseline shows fused+ACF+perm > 30% of per-series time after Z1/Z3 land; equal-length batches only; L·n·8 B ≤ ~16 KB. Plus the standing bars: ≥10% end-to-end gain for mainline inclusion, §9.4 stop rule.
  • Revive with: git show experiment/soa-4x:src/kernels/reduce.rs (lines 821–1153) and git show experiment/soa-4x:src/pipeline.rs (lines 252–260), then parity-test against NumPy references before wiring (§14 claims policy applies to the revival PR too).