merge: sync main with upstream/main

[code-yeongyu]

Summary

• Merges upstream/main at 3da591a through a history-preserving merge commit.
• Resolves catalog, native provider, llama, prompt propagation, Neo runtime, and compaction integration conflicts.
• Commits generated provider catalog data for reproducible offline builds.

Validation

• npm run check
• npm run build
• npm test
• Senpi CLI, RPC, mock-loop, TUI, and PTY smoke evidence under local-ignore/qa-evidence/20260719-upstream-sync/

Review

• Addressed independent review findings for prompt application, native provider ordering/startup, list-models discovery, sleeping runners, and deterministic tests.

---

Summary by cubic

Syncs main with upstream/main and moves the model catalog to committed JSON shards for reproducible, offline builds. Adds native provider registration and llama.cpp router support in the coding agent, fixes pricing and retry handling, and narrows catalog publish windows.

• New Features

  • Committed provider catalog JSON under packages/ai/src/providers/data/*; model modules now import JSON, build copies data, and generation is opt‑in via npm run generate:models and npm run diff:model-catalog (NodeNext module config and data-json.d.ts included).
  • Coding agent: register complete native providers and add llama.cpp router support (/login llama.cpp, /llama, Hugging Face search), with updated docs.
  • CI: limit publish-model-catalog.yml to 10:17/12:17/14:17 Europe/Vienna on weekdays and simplify SOURCE_REF resolution.

• Bug Fixes

  • Preserve GitHub Copilot long-context pricing tiers in generated metadata.
  • Show API-equivalent implied costs for Kimi Coding subscription models (tests updated to k2p7 and friends).
  • Classify OpenAI Responses streams that end before terminal events as retryable.
  • Fix prompt-template defaults for all arguments in the coding agent (${@:-default}, ${ARGUMENTS:-default}).

^{Written for commit 2532485. Summary will update on new commits.}