fix(agents): Ollama chat HITL — sanitize null content and preserve pending approval

[w7-mgfcode]

Problem

Two distinct defects break the agent-chat HITL approval flow on local Ollama (agent_default_model=ollama:*), discovered while dogfooding the Chat page.

Bug 1 — null-content crash (FallbackExceptionGroup)

Ollama's OpenAI-compatible /v1/chat/completions rejects any message whose content is JSON null and which carries no tool_calls with 400 invalid message content type: <nil> (stricter than the real OpenAI API). A weak local model can emit a degenerate empty assistant turn (no text, no tool call); PydanticAI serialises it as content: null and then replays that message on its validation-retry, so every retry 400s and the whole run dies with FallbackExceptionGroup (both primary + fallback). Not streaming-specific — hits agent.run() too (any multi-step/tool turn).

Verified truth table (curl → forecastlab-ollama, qwen3:8b):

• assistant content=null + tool_calls → 200 ✅
• assistant content=null no tool_calls → 400 ❌ invalid message content type: <nil>
• assistant content="" no tool_calls → 200 ✅

Bug 2 — pending approval dropped on model misbehavior

A gated tool (create_alias/archive_run/save_scenario) sets deps.pending_action the moment it fires (the HITL gate works), but it does NOT halt the run. A weak model rambles past the gate, calls more tools, and exhausts its retry budget → UnexpectedModelBehavior is raised before agent.run() returns. AgentService.stream_chat / chat's except UnexpectedModelBehavior handler then emits a generic "invalid tool call" error and returns without checking deps.pending_action — so the already-captured approval (with valid args) is silently discarded and the Approve card never appears.

Fix

1. Bug 1: inject a sanitizing httpx.AsyncClient into OllamaProvider(http_client=…) in build_agent_model (app/features/agents/agents/base.py) that coerces outgoing messages[*].content: null → "".
2. Bug 2: in the UnexpectedModelBehavior handlers of both stream_chat and chat, surface deps.pending_action (persist + emit approval_required / return pending_approval=True) before falling back to the generic error. The gate fired — the approval is valid regardless of post-gate model misbehavior.

Acceptance

• Outgoing Ollama chat requests never send content: null.
• When a gated tool fires but the model then misbehaves, the HITL Approve card is surfaced (not a generic error).
• Existing "invalid tool call" behavior preserved when NO gated tool fired.

[w7-mgfcode]

Fixed in 5fff334 (sanitize null content) + cc72f89 (preserve pending approval); shipped in v0.2.21. Closing.