refactor(api/nlq): convert UI actions from marker system to tool definitions

[SimplicityGuy]

Summary

• Converts the NLQ engine's 10 UI action types (seed_graph, switch_pane, filter_graph, ui_find_path, highlight_path, focus_node, show_credits, open_insight_tile, set_trend_range, suggest_followups) from a markdown HTML-comment marker (``) to real Anthropic tool-use definitions. Each action tool validates its payload via pydantic and records the Action into a per-request list that populates `NLQResult.actions`.
• Removes `_ACTIONS_INSTRUCTION`, `_ACTIONS_MARKER_RE`, and `_extract_actions` from `api/nlq/engine.py`. The system prompt now contains one short paragraph pointing the model at the action tools; the rest is self-documenting via Anthropic's tool schemas.
• `ui_find_path` is the tool name for the client action type `find_path` because the data tool `find_path` already owns that name in the Anthropic tool namespace. The internal mapping in `_ACTION_TOOL_TO_TYPE` rewrites it back to `find_path` when building the Action, so the client-visible wire format is unchanged.

Why

The marker system was load-bearing on LLM discipline and failed on complex queries in practice. Users asked things like "what's the biggest Trance label" and got a perfect markdown answer but zero actions fired — the model forgot the marker, or the regex silently dropped malformed JSON. Tool-use makes the action emission a structural contract: the model sees each action as a first-class tool with a JSON schema, and the engine assembles `NLQResult.actions` from the actual tool invocations instead of scraping text.

Contract preservation

• `NLQResult.actions` still ships as `[{type: str, ...payload}, ...]` on the wire — the NLQ router already serializes via `action.model_dump(by_alias=True, mode="json")`, and the Action pydantic models are unchanged.
• The frontend `NlqActionApplier` and handler shapes in `explore/static/js/nlq-handlers.js` and `explore/static/js/nlq-action-applier.js` are untouched — the tool input schemas match their expected payload shapes 1:1.
• Data tools are untouched. Auth-aware tool selection is untouched. `_AUTH_ADDENDUM` is untouched.

Test plan

• `uv run pytest tests/api/nlq/` — 110 passing (14 new action-tool tests plus existing)
• `uv run pytest tests/api/` — 1551 passing
• `uv run mypy api/nlq/` — clean
• `uv run ruff check api/nlq/ tests/api/nlq/` — clean
• Each of the 10 action tools exercised end-to-end (single-tool invocation → Action recorded in `NLQResult.actions`)
• Mixed data + action tool in the same iteration (search + seed_graph)
• No-action data-only flow leaves `actions` empty
• Invalid action payload (`switch_pane` with unknown pane) returns an error result and is NOT recorded
• Multiple distinct actions recorded in order (switch_pane + set_trend_range)
• Action tool names disjoint from data tool names

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 96.87500% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
api/nlq/tools.py	95.83%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[github-actions]

E2E Coverage (webkit)

Totals
Statements:	46.5% ( 1241 / 2669 )
Lines:	46.5% ( 1241 / 2669 )

[github-actions]

E2E Coverage (chromium)

Totals
Statements:	46.5% ( 1241 / 2669 )
Lines:	46.5% ( 1241 / 2669 )

[github-actions]

E2E Coverage (firefox)

Totals
Statements:	46.5% ( 1241 / 2669 )
Lines:	46.5% ( 1241 / 2669 )

[github-actions]

E2E Coverage (webkit - iPhone 15)

Totals
Statements:	46.5% ( 1241 / 2669 )
Lines:	46.5% ( 1241 / 2669 )

[github-actions]

E2E Coverage (webkit - iPad Pro 11)

Totals
Statements:	46.5% ( 1241 / 2669 )
Lines:	46.5% ( 1241 / 2669 )