Fix run_823.py phase-5 length-R2 diagnostic: valid_idx slice + un-swallow bare except

Overview / Motivation

Surfaced during task #823's pipeline (open concern phase5-length-r2-valid-idx-mismatch, raised by codex-code-reviewer r5, addressed for #823 by an inline analyzer recompute). The underlying code defect remains and violates the fail-fast hard rule.

Goal

Make the phase-5 length-R² diagnostic in src/explore_persona_space/experiments/issue_823/run_823.py fail loud and compute correctly when contexts are dropped: slice the span-length arrays by common_valid_idx before correlating, and remove the bare except Exception: logger.warning(...) that silently swallowed the pearsonr length-mismatch crash (~L2103).

Defect

validity_diagnostics.json.length_r2_correlation came out {} on the production run: phase 5 paired full-length (5000) span arrays with valid-row-only (4998) per-context R² arrays; scipy.stats.pearsonr raised; a bare except swallowed it. The pre-registered length-confound diagnostic was silently missing from the artifact (rescued post-hoc by the analyzer with span[common_valid_idx] alignment).

Fix sketch

• Persist/reload common_valid_idx in phase 5 and slice phase3_span_lengths by it before any correlation.
• Replace the bare except with a raise (or at minimum a hard error that fails the phase) — no silent defaults.
• Add a small regression test (a dropped-context fixture) asserting the diagnostic is non-empty and the crash is not swallowed.

Scope

src/explore_persona_space/experiments/issue_823/run_823.py (+ a test). Experiment-code hygiene for the #823/#779 line's future reuse; no re-run of #823 needed (its diagnostic was already rescued inline and is in the clean-result).

Provenance

• Raised on task #823 (concern phase5-length-r2-valid-idx-mismatch; addressed-for-823 record in concerns.jsonl 2026-07-03).