Harden chat, runtime, and generation stream reliability

[dovvnloading]

Summary

This QA pass fixes four high-confidence reliability issues across chat generation, the managed llama.cpp runtime, and the generated API contract.

• Regenerate responses from the selected turn's own prompt and attachments instead of component-wide state that can belong to another chat.
• Preserve next-turn text and attachments while a generation request is being accepted, including both attachment-staging race orderings and the temporary-to-persisted thread ID transition for new chats.
• Reap a llama.cpp server process whenever startup fails, escalating from terminate to kill when the process does not cooperate.
• Describe the generation event stream as text/event-stream in OpenAPI instead of incorrectly advertising JSON.

Root cause and impact

Chat regeneration reused mutable page-level prompt and attachment state. After navigation, that state could belong to a different thread, while delayed request acceptance could clear or strand drafts created for the next turn. The fix derives regeneration inputs from the selected transcript turn and gives each in-flight attachment batch a request-local draft target that can be safely retargeted before new-chat navigation.

The llama.cpp manager did not publish a process into managed state until it became healthy. A process that timed out during startup could therefore survive the failure path without being wait-reaped or kill-escalated, continuing to consume CPU, RAM, or GPU resources.

The generation events route returned SSE at runtime but its generated OpenAPI response still declared application/json, which could misconfigure generated clients and API tooling.

Validation

• ./scripts/check.ps1 -Tier full
  • Ruff: passed
  • Backend: 509 passed, 1 platform-specific skip
  • Generated contract drift: passed
  • Python compile check: passed
  • TypeScript typecheck: passed
  • ESLint: passed
  • Frontend unit tests: 161 passed
  • Playwright browser tests: 18 passed
  • Production frontend build: passed
• Repository pre-push quick gate: passed
• Focused regression coverage added for cross-chat regeneration, draft/attachment acceptance races in both orderings, new-chat scope migration, SSE media type, and uncooperative runtime shutdown.
• Final diff review: no outstanding findings.

Operational notes

No dependencies, persistence formats, or database migrations change. Rollback is a normal revert of this commit.