Add versioned benchmark configuration and recoverable Studio runs

[cgmattos]

Configured benchmark runs now resolve a frozen configuration revision, retain complete report evidence, and can pause or explicitly resume without replaying finalized attempts. Settings provides the versioned configuration editor; interrupted work keeps its original identity, evidence and spending ceiling.

Changes

• Add Git-backed configuration storage, validation, conflict-safe saves to main, and read-only snapshots when repository access is unavailable.
• Add configured Pause/Resume with a recovery preview, exclusive execution ownership, persisted cancellation, and billing/version/approval checks before continuation.
• Read reports from retained versioned attempt evidence and update Activity from configured result events, including recovered attempts.
• Include the Gmail front-door translation correction and operational recovery/deployment records. AutomationBench inputs and historical verdicts remain unchanged.
• Integrate current main (0d7a127), preserving its updated run page. Remove its stale test assertion for the already removed Actions column.

Validation

• Integrated Python regression group: 131 passed; the sole stale run-page assertion was corrected, then all 9 run-page tests passed.
• Real browser pause/drain/resume: 4 free attempts in 2 segments, matching uninterrupted execution and preserving earlier results.
• Configuration editor browser checks passed on desktop and 390px, including validation, conflict handling and draft retention.
• Hosted controls/report verification: 19 deployed source hashes matched the reviewed upload; 28 historical job/event/result files across 12 jobs remained byte-identical.
• No paid benchmark was launched. CI runs tests and corpus validation, with no deployment step.

Remaining prerequisites

Live Git saves still require the restricted PAT to gain repository access; hosted configuration stays read-only meanwhile. The pre-existing hosted world differs from the frozen tier-simple world, so version checks continue to block another launch until the matching upstream runtime is restored. The current-main integration in this PR has not been redeployed.

Detailed evidence: monarch-benchmark/docs/rounds/2026-09-11-studio-run-controls.md.