feat: Token budget aware fresh tail

[Bikkies]

This basically just adds a way for us to force compact older messages when we actually reach the configured compaction threshold, instead of the blunt "keep last 64 messages no matter what" behaviour that could lead to overflows when using models with lower context lengths. Now we TRY to keep the last 64 (or whatever you configure) messages, but if they really don't fit, we gotta compact.

freshTailCount protects the last N messages from compaction regardless of size. With the default of 64, a handful of large tool results in the tail can exceed the context window and there's no way to compact them out, even when context is about to overflow.

This is common in agentic workflows where tool results vary wildly in size. One message might be 50 tokens and the next one might be 30k.

This adds a token-count cap to the fresh tail using the existing contextThreshold config. Walks backward from the newest message, accumulating tokens. Once the budget (contextThreshold * contextWindow) is hit, older messages lose protection and become compactable regardless of how many you try to protect with freshTailCount

[Bikkies]

2 tests fail because they use token budgets small enough that the fresh tail exceeds the compaction threshold which is the scenario this PR addresses.