feat(config): caps — auto conversation budget, 200 pairs, 32K batch cap, 8K render cap, 16K local completion cap, 60 worker steps, 32K auto-context floor

[plombeer31]

What

Before: fixed caps sized for the previous generation of the loop — a 20-pair conversation limit, a 16,000-char batch cap, a 4,000-char render cap, an 8K local completion cap, 40 worker steps and a 16K auto-context floor.

After: the conversation budget is auto (32K fallback), conversationMaxPairs 200, the batch result cap 32,000 chars, the render cap 8,000, the local completion cap 16K, 60 worker steps, and the auto-context floor 32,768 so it always covers the fixed prompt plus the completion cap. Existing explicit values are migrated only when equal to the old default.

Why

The earlier caps were tuned around a 16–32K local context; with 131K contexts, water-filled batch results and line-boundary paging the old numbers re-created the truncation the previous PRs removed. One commit so the numbers are reviewed together.

How it was verified

• npm run lint clean
• npx vitest run --minWorkers=1 --maxWorkers=3 src/agent src/config src/llm/run-mode src/local-llm src/prompt src/session — 103 files / 1631 tests green (baseline failures: none)
• verified live: the bench-2 matrix ran on these caps; the step-half hand-back landed at 30 of 60 worker steps

Stacked on #436; merge in order.