corpus: v_script_summary multiplies parse capture counts through one-to-many joins

[mobileskyfi]

Problem

v_script_summary joins artifact_files, highlight_snapshots, and parseil_results before summing parse capture outcomes. These one-to-many joins multiply the parse rows. The distinct counters survive, but SUM(CASE WHEN p.ok ...) does not.

Reproduced on the pinned schema-v4 corpus:

• Repository revision: 7d623557d257e0a4d50f47c87d6a842ff2699401
• Database SHA-256: 4f1560f757265570b5b04c9f93d6d57b432064cf655ce2d1852da95536e313a9

Source	Actual successful capture rows	View's parseil_ok_count
edge-cases/empty.rsc	6	72
edge-cases/unicode-heavy.rsc	6	72
corpus row 228	6	72

edge-cases/oversize-32k.rsc has six failed capture rows, but the view reports parseil_error_count=36.

Reproduction

Run against that pinned database:

SELECT s.id, s.path, v.parseil_version_count,
       v.parseil_ok_count, v.parseil_error_count,
       (SELECT COUNT(*) FROM parseil_results p
        WHERE p.script_id=s.id AND p.ok=1) AS actual_capture_ok,
       (SELECT COUNT(*) FROM parseil_results p
        WHERE p.script_id=s.id AND p.ok=0) AS actual_capture_error
FROM v_script_summary v
JOIN source_scripts s ON s.id=v.id
WHERE s.id IN (3,4,6,228);

Source: scripts/build-corpus-db.ts, CREATE VIEW v_script_summary.

Scope and acceptance

• Aggregate capture outcomes per script before joining unrelated one-to-many tables, or use equivalent independent aggregates.
• Assert that adding an artifact or highlight snapshot cannot change parse outcome counts.
• Verify every source's view counts against direct queries of parseil_results, including failures and missing captures.
• Regenerate the database reproducibly and update downstream pins deliberately.
• Keep capture success distinct from script validity (corpus: separate parse outcomes from rejection causes and bind captures to exact inputs and device environments #37): ok=1 means capture succeeded, even when its response is a diagnostic.

This finding affects this summary view. It does not explain the Glossema comparison's direct-join counts; those have separate classification problems in #37.

Independent offline review by Codex, 2026-09-21; filed with maintainer authorization.