Fix Mean Citedness tab: outcome classification, denominators, caching

[LukasWallrich]

Part of a repo audit. The Mean Citedness tab disagreed with the rest of the dashboard about which studies count as successful/failed:

• Outcome classification: render_impact_factor.R used exact string equality (== "successful"), silently dropping variant labels ("success", "failure", "successful || success", "statistically successful but flawed", …) that the frontend's classifyOutcome() catches. The R script now uses a faithful port of that JS function, so all tabs agree. Effect on the committed data: n_success 921→970 (+49), n_failed 826→851 (+25).
• Percentage denominator: success/failure percentages divided by n_total, which includes ~20 unclassifiable rows ("na", "uninformative", "descriptive only"). The R script now emits n_classified and the frontend divides by it.
• Dead "Inconclusive" category: the token never occurs in the data, so the stat and histogram series were permanently zero; removed from R output and the chart (the classifier still recognizes the label should it ever appear).
• Tab reload churn: the tab re-fetched the JSON and re-rendered both Plotly charts on every visit; it now caches like the Authorship Overlap tab, and the GAM scatter jitter is computed once per load so points no longer jump on re-renders/theme toggles.
• run_fect.R (dormant): the t=−1 reference row is now built from an existing emfx row instead of rbinding a 5-column data.frame that errors on column mismatch.

data/impact_factor_data.json was regenerated locally with the fixed script (R + mgcv); CI regenerates weekly. node --check passes on app.js.

🤖 Generated with Claude Code

[LukasWallrich]

Codex review (gpt, read-only): REQUEST_CHANGES — the statistics were verified clean; two caching findings, both addressed:

1. Medium — cached tab visits still re-rendered both Plotly charts (the guard called renderMcCharts instead of returning; theme toggles already re-render via _rerenderAllCharts). ✅ Fixed: cached path now returns immediately.
2. Low — concurrent tab activations could issue duplicate fetches while the first request was in flight. ✅ Fixed: _mcLoading in-flight flag, reset in finally so a failed load can retry.

Codex verified everything else in detail: the R classifier is a faithful port (including not robust → other, robustness challenge → failed, exact replicated → successful, NA/empty → other); GAM subset and binary coding correct; regenerated counts reconcile (970+851+500 = 2,321 of 2,341; n_model 1,821 = 970+851); the histogram's 12 "missing" classified rows are the intended OMC≥20 axis crop; all JSON field names match the frontend reads; jitter stable across theme re-renders; and the run_fect.R reference-row fix is safe.

(Also fixed during self-review before codex ran: the original guard checked window._mcData but never assigned it.)

[LukasWallrich]

Codex re-verified the caching fixes: _mcData assigned after parse, cached/in-flight activations return immediately, _mcLoading always cleared in finally so failed loads stay retryable, and theme toggles still re-render via _rerenderAllCharts. FIXES_OK.