fix(agents): address agent smoke review

- Require bash 4.4+ up front, and match answer facts with jq regexes
  (Oniguruma, (?i) everywhere) instead of GNU-only grep -P.
- Unparseable agent output fails that run instead of aborting the
  matrix; check_run now also reports unparseable tool calls, which
  previously passed the tool check silently.
- Copy the Hermes model block with awk up to the next top-level key, so
  a final or differently ordered block is handled.
- --keep keeps run output but deletes the copied agent credentials.
- docs: full path to the tool golden.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs/SMOKE.md

@@ -26,7 +26,7 @@ user-visible output changed. Regenerate with
 and commit the golden diff with the reason.
 
 `crates/idx-mcp/tests/mcp.rs` drives `idx-mcp` over stdio JSON-RPC with the
-mock provider. `tests/golden/tools.json` there pins what agents see (tool
+mock provider. `crates/idx-mcp/tests/golden/tools.json` pins what agents see (tool
 names, descriptions, input schemas); regenerate it the same way after an
 intended change.
 
@@ -91,7 +91,8 @@ tests/agents/agent-smoke.sh --live --case shareholders --keep          # real da
 - Each agent gets an isolated config that registers only the local server;
   your own agent settings and MCP servers are not loaded or changed. Logins
   are copied from the agent's home (Codex `auth.json`, pi and Hermes auth
-  files) into a temp dir that is deleted afterwards unless `--keep`.
+  files) into a temp dir that is deleted afterwards; `--keep` keeps the run
+  output but still deletes the copied credentials.
 - Tools are auto-approved for the run (Claude `dontAsk` with
   `mcp__idx__*`, Codex `default_tools_approval_mode="approve"`); Claude and pi
   also get no built-in tools, so they can only answer through MCP.