fix(providers): retry the local probe after a startup miss

[cursor]

Bug and impact

If llama-server is not listening when ShellClaw starts, local_init sets s_local_unavailable and that flag stays set for the life of the process on the default chain (anthropic, openai, local). Later chats return Local inference server unreachable without contacting the server, even after the model has finished loading.

That disables the offline path:

• fallback_chain = ["local"] never recovers until the process is restarted.
• When cloud providers fail after startup, the local fallback is still skipped.

Router recovery cannot clear the flag in the default order. It only probes backends that sit before the active index, and local is last.

Root cause

local_chat treated the one-time init probe as permanent. provider_local_recovery_probe_succeeded() only runs from that earlier-index recovery scan, so a startup miss on the last chain entry is never retried.

Fix

While the flag is set, local_chat probes /v1/models or /health again. A successful probe clears the flag and the turn is sent. A failed probe still returns the unreachable error and the router can fall through.

Validation

• Added test_chat_reprobes_when_server_comes_up. It failed before the fix (chat stayed unreachable after init) and passes after.
• make test_local_provider
• make test_router