Alpha gate F: package three real-world case studies (bad → fixed → what others miss → how Own reports it)

[PhysShell]

Agent brief — tier: basic. One PR for all three studies (Closes #<this>); branch from main. House rules: AGENTS.md.

Goal

Gate F of docs/notes/alpha-readiness.md: the raw material for three case studies exists, the packaged studies don't. Write three one-page studies in the format bad → fixed → what others miss → how Own reports it.

The three studies (material already on record)

1. ScreenToGif VideoSource — the flagship view→view-model subscription leak.
2. ScreenToGif SystemEvents leaks (×2) — incl. the corpus/wpf/systemevents-region-escape reduction (OWN014 / region escape).
3. The Dispose-leak cross-tool Agree case — where Own.NET and CodeQL's dedicated query agree on the fixture (dispose-not-called-on-throw).

Sources of truth (cite, do not re-run or invent)

• docs/notes/real-world-mining.md — the findings, reduced code, and honest framing.
• docs/notes/oracle.md (+ oracle-known-fps.md) — every cross-tool claim (CodeQL / Infer# / NetAnalyzers) must trace to what is recorded there.
• corpus/ reductions where they exist — use them to show Own's actual diagnostic output.

Deliverable

• docs/case-studies/<slug>.md × 3, each: the bad code (reduced), the fix, what CodeQL/Infer# say (recorded results only), Own's actual report (diagnostic code + evidence).
• Single-line links added to the README "Case studies" placeholder section (do not restructure README — the front-door issue owns its structure).

Guardrails

• Honest claims only: "CodeQL/Infer# have no event-subscription query, so they stay silent" is fine; "we would have prevented X" is banned.
• Reproduce Own's output via the corpus reductions (python -m ownlang ... / scripts/own-check.sh) rather than paraphrasing it.

Done when

Three pages exist, each self-contained and honest, linked from README; ruff check . clean.