PR-CS-8: Association strength (ΔR²) — per-scope effect-size guide for the Analyze Wall

[jukka-matti]

What

Surfaces each factor's semipartial R² ("association strength, ΔR²") in the Analyze Wall's ModelBuilderBand — the unique share of the spread each factor accounts for inside the drilled scope. An association magnitude paired with the existing partial p, framed explicitly as a clue to investigate, never a cause verdict.

Closes PR-CS-8 of the connective-surface redesign (sub-plan: docs/superpowers/plans/2026-06-02-cs-8-association-strength.md).

Commits (8)

1. feat(stats) — perFactorDeltaR2 pure helper (semipartial R², O(1) reads off the enumerated subset index, no new regression; raw R² so ≥0, clamped)
2. test(stats) — empty-baseline ΔR² coverage + decouple tests from the subset-key encoding
3. feat(i18n) — 4 association-strength/ΔR² labels across all 32 locales + the MessageCatalog interface
4. feat(analyze) — render per-factor ΔR² bars + value + a "not a verdict" framing line; re-point captured topFactor to highest-ΔR²
5. refactor(analyze) — extract byDeltaR2Desc (single comparator for the 3 sort sites)
6. test(analyze) — lock in per-scope vital-few re-rank on drill + ΔR² numeric-signal guard
7. test(analyze) — make the re-rank test load-bearing (junk Noise candidate + balanced wobble)
8. docs(cs-8) — refine LOCKED chore(deps): bump actions/upload-artifact from 4.6.2 to 7.0.0 #2 → effect-size-with-p (spec §2 + decision-log + source headers)

Design (pinned in the sub-plan)

• Magnitude = per-factor semipartial R² (R²(kept) − R²(kept∖f) for a kept factor; gain-on-add for a candidate) — the same numerator the nested-F partial p is built on. The effect size paired with the existing p.
• Non-summing per ADR-073 — raw semipartials; correlated factors' shared variance is attributed to no single factor (the caption says so). No forced variance decomposition.
• Refines LOCKED chore(deps): bump actions/upload-artifact from 4.6.2 to 7.0.0 #2, does not reopen it: ΔR² is an effect size, not a model-selection criterion → still no Cp/BIC on the surface.

Grounding corrections during build (logged for the method record)

• Task 2 (i18n) was a 33-file change, not 1 — the plan assumed a derived MessageKey union, but MessageCatalog is a closed interface with 32 hand-maintained complete locale catalogs (English placeholders are the convention for technical labels, per fe-1/fe-2a/fe-2b). Synced via a deterministic script; verified by the tsc build + the 70-test i18n completeness suite.
• The re-rank integration test was made load-bearing — the first cut only had one eligible factor in the drilled scope (couldn't tell re-rank from "show the sole factor"). Added a junk Noise candidate that must NOT be kept in Region A, so the assertion genuinely requires per-scope recomputation.

Verification

• pr-ready-check.sh: ✓ all checks passed (full turbo test + 4 builds + lint + boundary + dist integrity).
• Final adversarial review (Opus): Ready to merge — Yes, no Critical/Important. Independently re-ran both full suites (core 3507, ui 2861), all 4 builds, and a standalone statistical probe of the fixtures; confirmed raw-rSquared correctness, ADR-073 non-summing, never-a-verdict framing, i18n completeness, and that topFactor's two downstream consumers only map it to linkedFactor (no p-ordering assumption).

Non-blocking follow-ups (Minor, from final review)

• Theoretical edge: if a multi-factor reduced OLS subset were ever skipped during enumeration, the kept-branch fallback would over-state ΔR². Near-impossible in practice (the displayed kept set is always enumerated; a reduced set has strictly fewer columns). Left as-is.
• Narrow-band: the kept row packs DeltaBar + ΔR² + p next to the factor name; the container is overflow-auto so it scrolls, not clips. Worth a glance at laptop width in the real Wall.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
mean-beoynd-lite-pwa	Building	Preview, Comment	Jun 3, 2026 11:18am
variscout_website	Building	Preview, Comment	Jun 3, 2026 11:18am