CL-6204: fire budgeted compaction, detect context overflow before send

[TheGreatAxios]

Summary

• Adds a per-model context-window budget (context-budget.ts), sized from InferenceSource.quirks via the same numCtx shape @corbits/ollama-adapter resolves for a live request, falling back to a conservative constant for sources with no quirks.
• Adds a summarize-budgeted-turns compactor that folds history against that budget instead of a fixed turn count (deliberately lossy — a bounded recap of the active window, not a durable record).
• WorkbenchDirector now actually fires it, at the one point in the reactor's event flow where caps.compact cannot stall a reply: a non-final tool.done inside a multi-call batch. @intx/agent's createAgent only threads env.compactors into the reactor (no ContextTransform seam exists at that layer — verified against agent.js), and compact cannot pair with infer in the same cycle with no follow-up event, so every other cycle risks stalling a reply if it compacts instead of inferring; compaction there is deferred to the next safe point.
• Before any decision that would infer, the director checks the budget's hard limit. Past it, it replies with the same "exceeded the model's context limit" message hosted providers produce via inference.error's context_overflow category, instead of sending a request Ollama would silently truncate (HTTP 200, no error).

Why

@intx/inference's reactor resends the full turn history every inference cycle. Hosted providers fail loud on overflow (context_overflow classification → friendly reply); Ollama returns 200 and truncates server-side, so a long local-model conversation degrades into incoherence with no signal. The existing summarize-older-turns compactor was registered but never invoked by any director (CL-6204's original gap).

Scope note

The owner's broader design for this ticket also calls for: two distinct compactor strategies (conversational vs. orchestration, the latter tracking cross-thread/agent status rather than prose), and a compaction-time sweep that extracts durable facts to @corbits/memory-tools before discarding older turns (an extra inference call per compaction). Both are real, larger pieces of work — the memory-sweep needs an inference handle threaded into the compactor seam (ContextStrategy.apply doesn't have one today), and the orchestration variant needs the whole-chat-manager surface identified and scoped. Neither is half-wired here; this PR ships the budget/overflow foundation both would sit on top of.

Confirmed while reviewing: the assistant workflow (workflows/assistant) already pins @corbits/memory-tools and its system prompt directs the model to memory_search/memory_add/memory_list for anything worth recalling — the chat host (packages/chat's platform-adapter.ts, via @corbits/agent-lifecycle) runs through this same workflow, so retrieval is reachable today for the channel-agent surface. It is on-demand only (no automatic per-turn injection), matching the owner's stated design.

Test plan

• bun test in apps/sidecar — 38 pass, 0 fail
• tsc --noEmit in apps/sidecar — clean
• Red/green covered: a short conversation stays untouched; a conversation past budget folds rather than resending whole; a minimum verbatim tail always survives; a prompt whose history exceeds the hard limit gets an honest reply instead of an infer that would silently truncate