NodeSlide durable sessions, hard budgets, and benchmark gates

[HomenShum]

What changed

• adds a durable NodeSlide session/job journal with lease-fenced retries, cancellation, review finalization, and append-only model/web receipts
• enforces hard per-run provider budgets before dispatch, including durable reservation, settlement, unreconciled exposure, and replay-safe accounting
• binds proposal acceptance to the exact durable job, candidate, deck, owner, and review state
• adds fixed-corpus UXBench/TasteBench schemas, deterministic PR gates, run/reward records, and agent-operability checks
• archives and consolidates the prior parallel work into one clean integration branch

Why

Long-running agent work must survive retries without duplicate model charges, expose honest cost and state, and keep all deck mutations behind human review. QA also needs a fixed corpus and reproducible evidence contracts instead of improvised prompts.

User impact

NodeSlide runs now have an authoritative durable lifecycle, hard budget enforcement, replay-safe provider calls, and review receipts that cannot be accepted after a failed or stale job.

Validation

• pnpm typecheck
• pnpm exec tsc -p convex/tsconfig.json --noEmit
• pnpm test — 133 files / 1,010 tests
• pnpm build
• pnpm qa:manifest
• pnpm test:nodeslide-bench — 20/20
• pnpm nodeslide:bench:pr
• pnpm lint:nodeslide-agent-ui — 9/9

Live fixed-corpus evidence and production deployment are tracked as the final merge gate; no improvised prompts will be used.