refactor: move durable compaction engine into core

[MyPrototypeWhat]

What

Moves the durable-compaction engine — planCompaction / compactHistory — out of @context-chef/ai-sdk-middleware and into @context-chef/core, where it belongs: the algorithm operates on the provider-agnostic IR (Message[]) and only ever touched AI SDK types at the boundary. The middleware functions become thin wrappers.

This corrects an altitude mistake from a461ef8 (unreleased, sitting in release PR #38): the library already factored its compaction primitives (groupIntoTurns, summarizeHistory, compactMessages) into core, and middleware's summarizeMessages was already a thin fromAISDK → core → toAISDK shell. The two new helpers broke that contract by putting the engine in the boundary layer.

Doing this before #38 merges means core ships the helpers correctly from day one; the middleware public API is never observed with the logic in the wrong layer.

How

• @context-chef/core gains planCompaction(history, options) and compactHistory(history, compress, options) over Message[].
  • compactHistory takes an injected compress callback (NOT a model), same pattern as summarizeHistory, so core stays provider-agnostic — no ai / @ai-sdk/provider dependency.
  • Returns the input history reference unchanged on a no-op (nothing old enough, or a blank summary), so callers can skip persistence via result === history.
  • Input contract (documented): a flat Message[] with system inline; direct adapter users (fromAnthropic/fromOpenAI/fromGemini) reassemble [...system, ...history].
• @context-chef/ai-sdk-middleware planCompaction / compactHistory become thin AI-SDK wrappers (fromAISDK → core → toAISDK), binding model via createCompressionAdapter. Signatures and behavior unchanged, including the no-op reference guarantee (result === ir ? prompt : toAISDK(result)). createCompressionAdapter is exported module-level (not in the package public index).
• generateText stays in middleware — the only true AI SDK binding.

Semver

@context-chef/core: minor (new additive exports). @context-chef/ai-sdk-middleware: patch (internal refactor, API unchanged). @context-chef/tanstack-ai auto-patches via updateInternalDependencies (dependency-range update only).

Verification

• pnpm -r typecheck clean (3 packages)
• Tests: core 714, ai-sdk-middleware 91, tanstack-ai 66 — all pass
• pnpm lint (biome) clean (100 files)
• Behavior parity proven by keeping the original middleware test suite green through the refactor; then slimmed (algorithm assertions migrated to core's new durableCompaction.test.ts).
• Reviewed adversarially: no behavior defects found (no-op reference parity, byte-identical AI-SDK output, no AI-SDK leak into core, no import cycle, semver all confirmed). Doc-precision and edge-case test gaps surfaced by the review were fixed in this branch.

🤖 Generated with Claude Code