Preset flag + self-describing run manifest in release report/leaderboard export

[ElbertePlinio]

MVP-completion slice: stop selecting official tasks by hand and make every published run self-describing.

• --preset flag in the headless CLI selecting active admitted tasks by metadata (single flag over task.yaml metadata — no preset registry; there is one real preset today)
• Release report + leaderboard export name the preset, exact task ids, task versions, and a corpus manifest digest (pin published numbers to a frozen corpus snapshot, not a drifting HEAD)
• Export raw trial rows and first-class f2p/p2p fields (baselineHiddenFailed / referencePublicPassed / referenceHiddenPassed) — fields only; no dashboards until the corpus is bigger
• Record per run: harness version, prompt template, tool permissions, timeouts, retry policy, model/provider config

[ElbertePlinio]

Parent plan: #10

[ElbertePlinio]

Started on feat/preset-manifest: --preset mvp metadata selection, corpus manifest digest, first-class f2p/p2p + raw trials, full run provenance in exports.

[ElbertePlinio]

Shipped via PR #29 (squash-merged, CI green, Codex review closed at round cap). Deeper per-entry corpus-manifest validation deferred to #6.