[wolfy-j]

Purpose

The agent node's checkpoint threshold compared the last action's prompt_tokens. With prompt caching (the history_tail marker from #90), prompt_tokens counts only the uncached suffix of the prompt, so a long conversation whose prefix is cached stays far below the threshold and never checkpoints, until the model call fails on the context window. The session host moved to context_tokens for the same reason.

Change

• The threshold compares tokens.context_tokens (uncached input plus cache reads and writes), which the llm library sets on every normalized response since 0.5.1 (the module already requires >= 0.5.3).
• The checkpoint binding refs, the function fallback payload, and the checkpoint marker metadata carry context_tokens / checkpoint_at_context_tokens next to the unchanged prompt_tokens / checkpoint_at_prompt_tokens, so custom summarizers keep their inputs.
• The default summarizer prompt states context_tokens next to prompt_tokens.
• The recovery stub model can report cache_read_tokens (scenario input) to exercise cached prompts; the stub summarizer records the context_tokens it receives.

Tests

• New threshold counts cached prompt tokens: 200 uncached + 5000 cached tokens against a 1000-token threshold must checkpoint and record 5200 context / 200 prompt tokens. It fails on master (no checkpoint).
• The function-fallback payload's context_tokens is asserted (5200 in the cached case, 500 in the uncached threshold case). The binding path builds refs from the same values; the suite has no successful checkpoint binding stub to observe it.
• Full suite (make test-sqlite) passes, lint clean.

[wolfy-j]

Hold until #136 is in: this makes checkpoints fire on cached agents, and on master a checkpoint leaves tool results without their calls. #136 fixes that.