get_registered_tool_map() rebuilt on every label resolution

resolve_tool_display_name is called in two hot paths: once per action in _humanise_message and once per streaming tool-start/tool-end event in reasoning.py. Each call re-constructs the full dict[str, RegisteredTool] from the cached snapshot. The snapshot itself is @lru_cached, but the dict rebuild is not.

For interactive streaming with many tool events this is redundant allocation on every event. Iterating the cached snapshot directly avoids the extra dict construction:

def resolve_tool_display_name(tool_name: str) -> str:
    """Return a human-friendly label for a tool name."""
    snapshot = _load_registry_snapshot()
    tool = next((t for t in snapshot if t.name == tool_name), None)
    if tool is not None:
        return tool.display_name or tool.name.replace("_", " ")
    return tool_name.replace("_", " ")

Three tools that only lived in the old TOOL_DISPLAY dict (query_kubernetes_logs → "Kubernetes logs", query_elasticsearch → "Elasticsearch", get_deployment_status → "deployment status") were never annotated with display_name. They now fall back to de-snaked lowercase strings like "query kubernetes logs" instead of "Kubernetes logs". The PR description says existing labels are preserved, but these three regress silently.

I checked the current registered tool names after the merge and those exact identifiers (query_kubernetes_logs, query_elasticsearch, get_deployment_status) do not appear to be live registered tools on this branch. For example, Elasticsearch is currently registered as query_elasticsearch_logs. I left them out to avoid preserving stale aliases, but happy to add a small legacy fallback if we still expect those names to appear in streaming events.

@imjohnzakkam yes a fallback makes sense right?

@muddlebee will do that and update

@muddlebee Added the small legacy fallback in 314535f for the three stale event names:

• query_kubernetes_logs -> Kubernetes logs
• query_elasticsearch -> Elasticsearch
• get_deployment_status -> deployment status

Also added a regression test to preserve those labels. The follow-up format-only commit is 96f4c12, and local checks are green:
ruff check app/ tests/, ruff format --check app/ tests/, and the focused pytest set all pass.

display_name is None shouldn't be a registration trigger. The any-arg check is meant to distinguish bare @tool from @tool(...). display_name is pure cosmetic metadata — adding it here means @tool(display_name="X") with no name/source would attempt registration and fail validation rather than being treated as a bare decorator. Recommend removing display_name from this check.

Fixed in b88019a by removing display_name from the registration trigger check.

Fallback returns tool_name.replace('_', ' ') without .title(), so unknown tools render as "my custom tool" rather than "My Custom Tool". The test even asserts the lowercase form. Intentional? The old reasoning.py fallback was also lowercase, so this is consistent — just flagging since display_name-annotated tools will be Title Case and fallback tools won't be.

Intentional for now. This preserves the previous fallback behavior from reasoning.py; registered tools can opt into title-cased labels via display_name.

The _require_non_empty_strings validator now runs on display_name too. That's correct. One edge case: an empty string display_name="" will raise a ValueError at registration time with a message referencing the field name. Consider ensuring the error message is clear enough to debug (Pydantic usually includes the field name, so this is probably fine).

Keeping the shared validator here. Pydantic includes the field name in the validation error, so the failure should still point clearly at display_name.

test_resolve_tool_display_name_falls_back_for_unknown_tools calls resolve_tool_display_name directly without any monkeypatching. This hits the real registry cache (_load_registry_tool_map). If a real tool named my_custom_tool is ever added to the repo, this test will silently start returning its display_name instead of the fallback. Mock the registry or use a name that is guaranteed to not collide (e.g. a UUID-ish sentinel like nonexistent_tool_xyz).

Fixed in b88019a; the fallback test now uses a sentinel name (nonexistent_tool_xyz_sentinel) so it should not collide with future real tools.