docs: assessment of the 2009 design notebooks (June, July, September)

[kgdunn]

Records what in the June, July, and September 2009 design notebooks (the earlier workflow-DAG chemometrics tool) is superseded by ScorePilot's current architecture, and what is worth building on.

Superseded: lazy/copy-on-change workflow copies, traits/weakrefs, custom digraph, coroutine pipelines, masked-array exclusion bookkeeping — all dissolved by immutable datasets + per-variant PreprocessingSpec + parent_id lineage.

Worth building on (ranked in the doc):

1. Score excluded samples against the fitted model ("move to validation rather than exclude") and record it in the Logbook
2. Named validation/test sets + an apply-model-to-new-data endpoint (currently no out-of-sample mechanism exists)
3. Model inversion for product/process design (Jaeckle & MacGregor): invert PLS/PCR from a desired Y spec to candidate X conditions, with Y-space feasibility pre-check and null-space exploration
4. "Transformations never alter shape; tools may" invariant for future lags/batch-unfolding support
5. Preprocessing breadth: MSC/SNV, smoothing, trimming, winsorising; robust PCA
6. Monitoring triggers/notifiers over T²/SPE limits (future module)
7. Adaptive model update ("refresh this model on dataset D" as a child variant)
8. Multiblock / SIMCA / PLS-DA / missing-data-tolerant fitting, upstream in process_improve
9. UX: variable-range name patterns, odd/even row splits, scatter-plot matrix / correlation map / sparklines

Docs-only change; version bumped 0.22.0 → 0.22.1 with uv.lock refreshed. (The second July upload was byte-identical to the notebook already covered.)

https://claude.ai/code/session_01CErviAK6p4MuZiVWsQT3vR