fix(server): webhook reviews fall back to hot-applied server LLM configs (0.10.1)

[Liewzheng]

Root cause

The webhook review path (run_review_common in src/server/mod.rs, shared by the GitLab handler src/server/gitlab/hooks.rs and the GitHub handler src/server/github.rs) resolved LLM providers from only config::resolve_config(None) (the config file [[llm]] section) and llm_configs_from_env() (LLM_CONFIG). It never read the hot-applied state.llm_configs (populated from the DB llm_providers table at startup and updated live by the WebUI / PUT /api/v1/config).

Result: providers configured in the WebUI had no effect on webhook-dispatched reviews — e.g. NAS deployments where every MR review failed with all experts failed: ... LLM config 'default' has no api_base set.

Fix

• run_review_common gains a trailing server_llm_configs: Option<Vec<LLMConfig>> parameter; resolution extracted into a pure helper resolve_webhook_llm_configs with precedence config file [[llm]] > server (hot-applied) > LLM_CONFIG env. An empty server vec counts as "not provided" (same semantics as the 0.9.1 API-path fix); None (CLI/legacy, no AppState) keeps the exact pre-0.10.1 behavior.
• Both webhook handlers snapshot state.llm_configs (via the existing Weak<AppState> upgrade pattern) and thread it down to the review task. The GitHub handler previously kept only the task store from with_app_state; it now also holds a Weak<AppState>, mirroring the GitLab handler. The snapshot is taken lazily inside the review-dispatching arms (GitLab: Merge Request / Note / System Hook; GitHub: pull_request / issue_comment) so ping/push/ignored events pay zero cost.
• When all three sources are empty, a clear ERROR is logged before the task fails, with caller-appropriate guidance: server-state callers are pointed at the WebUI (Configuration → LLM), stateless callers at the config file / LLM_CONFIG env. Log-only — failure semantics unchanged.

Tests

• 4 new unit tests lock the precedence: file wins over server; server used when the file is empty; Some(vec![]) falls through to env; None keeps config-file → env behavior.
• New integration test llm_configs::webhook_review_uses_hot_applied_server_llm_configs (tests/server/llm_configs.rs) reproduces the reported scenario end-to-end: real spawned server with no LLM_CONFIG env and no config file, provider added via POST /api/v1/llm/providers (the WebUI flow) pointing at a wiremock LLM, a GitLab MR webhook fired against a wiremock GitLab instance. Asserts the webhook-dispatched review settles completed and the hot-applied provider was actually called (/chat/completions hits ≥ 1). Pre-fix this task failed with "no api_base".

Verification (run in the fix worktree)

• cargo fmt --check — clean
• cargo clippy --all-targets --all-features — 0 warnings
• cargo test — all green: lib 1525 passed / 0 failed (2 ignored), bin 58, cli 31, server integration 49 (incl. the new webhook test), doc 4

Version

Cargo.toml bumped to 0.10.1; CHANGELOG.md gains [0.10.1] - 2026-09-07. No CLI behavior change; no API/schema change.