v0.1.2

Changes

• Resolve compacted corpus refs into stock OpenAI and Codex agent prompts, with an 8 KiB default prompt ref budget and WithPromptRefBytes(max) tuning.
• Add Compactor.Corpus() so LLM receivers can resolve refs from the same corpus used for compaction.
• Preserve shutdown cancellation in dead-letter dispatching and make dead-letter enqueue honor already-canceled contexts.
• Make MemCorpus.Get and MemCorpus.Merge honor canceled/expired contexts without mutating on canceled merge.
• Make budget exhaustion and Done() closure linearizable for observers.
• Sync scaffold runtime references.
• Bump HandoffKit plugin manifest to 0.1.2 and retarget the plugin description toward coding agents.

Verification

• go test -count=1 ./llm
• go test -count=1 ./runtime
• make test
• make build
• make vet
• make check-plugin-sync
• make test-integration