ga: diagnose blank LLM responses; tui keeps [Warn]/[Error] labels

[louisss1016]

Addresses #815.

What

Blank responses — observed with third-party relays that return an HTTP-successful but whitespace-only payload — used to produce a single opaque line with nothing to attribute the failure to. This PR adds a redacted diagnostic and keeps severity visible in the TUI, without touching the retry semantics from #201.

Changes

ga.py

• describe_blank_response() — pure, module-level helper producing a compact one-liner: whitespace_only vs empty_payload classification, len(content)/len(thinking), deduped+sorted response block types (extracted from response.raw), and, when a backend is reachable, model= and host=.
• _url_host() — hostname-only extraction (no userinfo, no port, no credentials); returns unknown for garbage input instead of raising, since a diagnostic must never crash the retry path.
• do_no_tool() — retry logic untouched (still 3 attempts, then exit). The warning line now carries the attempt counter and the diagnostic:
  [Warn] Empty LLM response (1/3); retrying [kind=whitespace_only len(content)=1 len(thinking)=0 blocks=text model=gpt-4o host=api.openai.com]

frontends/tui_v3.py

• _ACTION_RE split into _ACTION_RE (Action|Status|Info|Debug) + _SEVERITY_RE (Warn|Warning|Error). Non-severity tags still collapse to · , but severity tags now render as · [Warn] … — issue suggestion timeout to any question #3.

Tests

• frontends/tests/test_blank_response_diagnostics.py — 15 tests following the repo's ast/exec extraction pattern (ga.py/tui_v3.py have heavy import side effects): classification, credential redaction (user:s3cret@ never appears), host edge cases, regex disjointness, and the exact rendering contract from the issue.

Verification

• python -m pytest frontends/tests -q → 287 passed; the 3 failures (test_data_backup symlink ×2, test_release_qualification) are pre-existing on Windows without developer mode / without a built package — identical before and after this change.
• Diff verified against a fresh tarball of f308ee7e: only the 4 hunks above plus the new test file.

Scope notes

• The lower-level llmcore.py path (ConnectionError("empty response") when a stream yields nothing) is already distinguishable from this agent-level guard and is intentionally left as-is.
• Issue suggestion 无法操作微信，也无法操作浏览器，操作微信需要做什么配置？ #5 (docs about reproducing against the official endpoint before filing relay-specific issues) is a docs decision — left for maintainers.