llamastash / llamastash / 26751491043

push

github

fix(daemon,proxy): port collisions + embedding auto-start

Three related fixes that all surfaced from a single failing
`nomic-embed-text-v1.5.Q2_K` launch:

- Proxy auto-start: propagate the GGUF `mode_hint` into the proxy's
  `CatalogRow` and through `StartParams.mode` so embedding / rerank
  models get the right `--embeddings` / `--rerank` argv. Before this,
  every model the proxy auto-started defaulted to chat mode and any
  `POST /v1/embeddings` against an embedding-only model returned 501.

- Port allocator: `reserve_port`'s explicit / `prefer_port` branch
  now probes via `try_bind_probe` before reserving, so a stale
  last-used port held by an external `llama-server` no longer rides
  through to llama.cpp's bind and crashes the launch. New
  `try_bind_probe` is a thin public alias of the existing internal
  probe.

- Orphan tracking (Tier A of the adoption story): supervisor stamps
  `LLAMASTASH_LAUNCHED=1` on every spawned `llama-server`. The sweep
  reads `/proc/<pid>/environ` (new `with_environ` refresh) plus
  `--port` / `-p` / `--port=N` from argv via the new `extract_port`
  helper, exposing both on `ExternalProcess`. `collect_in_use_ports`
  extends to include marker'd external ports so the allocator skips
  them proactively, and `status --json` surfaces the new fields on
  each external row.

159 of 190 new or added lines in 11 files covered. (83.68%)

42169 of 48849 relevant lines covered (86.33%)

1633.46 hits per line