refactor(server): replace custom compaction with AI SDK guided pruning

Collapse the four-stage compaction pipeline (prune, tool-output reduction,
LLM summarization, sliding-window fallback) into the shape the AI SDK
compaction guide recommends: one trigger threshold, one prune pass, one
deterministic prune-only fallback.

The sliding-window fallback is replaced by pruneMessages hard mode at all
four of its former call sites, which cannot orphan a tool call from its
result. LLM summarization and the split-turn dual-summary path are removed.

Token estimation stays image-aware rather than adopting the guide's
JSON.stringify heuristic, which would read a single base64 screenshot as
roughly 250K tokens.

Fixes three latent issues along the way: the chars-per-token constant was 3
in the estimator and 4 in the inflation guard, the fallback targeted a
looser budget than the trigger so a compaction could succeed and re-fire
immediately, and a zero or negative client-supplied context window produced
a threshold every request exceeded.

Author: DaniAkash