[c-mongan]

Source inspection previously required repeated packet submissions even when the user only needed a few verified findings. Add a compact reading view and one-shot finish command that saves selected observations through bounded leases, while preserving originals and legacy interfaces.

Changes:

• Derived rolling-caption cleanup and explicitly labelled lexical search fallback.
• Idempotent selected inspection with unchanged five-minute imports and no automatic inspection credit.
• Acquisition and engine gaps bound into review views and retained in finish receipts; independent review regression repaired.
• Public repeated comparison and honest experimental recommendation.

Validation:

• 98 local tests: 96 pass, 2 optional skips; both live converter tests passed separately.
• Fresh standalone skill copy: real PPTX prepare/read/finish/retry/query/export passed from outside checkout.
• Skill validator and diff checks pass.
• Two video cases, two repetitions: both conditions 12/12; SourcePack 358.82 seconds versus direct 241.51 seconds (49% slower). Five document cases and fresh-context frame/policy retrieval worked in both conditions.

Limitations: two SourcePack timestamps identify broad transcript spans. A policy paraphrase required query rewording. No general accuracy, speed, semantic retrieval or adoption advantage is established. Raw evidence, benchmark keys and machine configuration remain private. The measured frozen implementation predates the separately tested gap-reporting repair.