chore(lint): auto-fix 253 of 283 ruff issues (F401, I001, E402, RUF100, UP037, SIM105)

Mass ruff auto-fix:
  - ruff check --fix: 109 issues fixed (F401 unused imports,
    I001 unsorted imports, UP037 quoted annotations, SIM105
    suppressible exception, RUF100 unused-noqa)
  - ruff check --fix --unsafe-fixes: 22 additional issues
  - ruff format: 70 files reformatted
  - Manual pass: fix 16 misplaced import httpx lines
  - Manual pass: fix remaining E402 (import-after-docstring)

Result: 283 errors -> 30 errors.

The remaining 30 are real issues that need manual review:
  5 F401 unused-import (likely auto-generated stubs)
  5 F821 undefined-name (real bugs in code that references
    redis/pydantic/LLMRegistry without imports)
  3 BLE001 (the compliance LLM fallback is intentional; the
    other two are real)
  3 RUF012 mutable-class-default
  3 SIM105, 3 SIM117, 2 E722, 2 E741
  1 B007, 1 B025, 1 E402, 1 RUF200 (pyproject.toml issue)

Tests: 436/437 pass (1 pre-existing SSE sandbox failure).
format check + import sort: now clean.
make ci: still gated on the 30 remaining real issues.
Follow-up: triage the 30 issues file-by-file.

templates/test_live.py

@@ -128,7 +128,7 @@ TEMPLATES = [
 async def test_all():
     failures = []
     async with httpx.AsyncClient(timeout=30) as c:
-        for tid, site, cat in TEMPLATES:
+        for tid, site, _cat in TEMPLATES:
             print(f"Testing {tid}...", end=" ", flush=True)
             try:
                 r = await c.post(
@@ -146,7 +146,7 @@ async def test_all():
                 print(f"ERR: {e}")
                 failures.append(tid)
 
-    print(f"\n{len([t for t in TEMPLATES])} tested, {len(failures)} failed")
+    print(f"\n{len(list(TEMPLATES))} tested, {len(failures)} failed")
     if failures:
         print("Failed:", ", ".join(failures))
     return 1 if failures else 0