Fixes #2197.

Root cause

attempt_self_reflection was called inside the per-tool result loop in
handle_native_tool_calls. When any tool returned an error (including permanent
errors like HTTP 403), the old code called attempt_self_reflection mid-loop,
which pushed User{reflection_prompt} + Assistant{response} into history
before the User{ToolResults} message was assembled. The resulting message
sequence sent to the API was:

[N]   Assistant{ToolUse(id-1, id-2, ...)}
[N+1] User{"You attempted to help..."}      ← reflection prompt (wrong position)
[N+2] Assistant{reflection response}
[N+3] User{ToolResult(id-1), ToolResult(id-2)} ← too late

OpenAI requires every tool_call_id in an assistant message to be immediately
followed by role:"tool" messages. Having a regular role:"user" message at
[N+1] → HTTP 400. Claude's context assembly downgrades orphaned ToolUse to text,
silently losing tool results.

Secondary issue: record_anomaly_outcome().await? propagated channel send errors
via ?, abandoning mid-batch ToolResult assembly before result_parts.push().

Fix

• Deferred self-reflection: removed attempt_self_reflection from inside the
  per-tool loop. A single pending_reflection: Option<String> captures the first
  eligible error output. After the loop assembles all ToolResult parts and
  user_msg is pushed to history, reflection fires — in the correct position
  after User{ToolResults}.
• Reflection errors swallowed: Ok(_) | Err(_) => {} — whether reflection
  succeeds, declines, or errors, Ok(()) is returned. ToolResults are already
  committed; the tool loop continues normally.
• record_anomaly_outcome errors ignored: changed await? to let _ = ...await so channel failures cannot abandon ToolResult assembly.

Tests

• R-NTP-13 (new): single ToolError::Execution(io::Error::other("HTTP 403"))
  → ToolResult present in history
• R-NTP-14 (new): two parallel permanent errors → both ToolResult parts present
• R-NTP-10/11/12 updated: now assert is_ok() (new contract — reflection
  errors no longer propagate) and search all messages for ToolResult parts

Checklist

• cargo +nightly fmt --check — clean
• cargo clippy --workspace -- -D warnings — clean
• cargo nextest run --workspace --lib --bins — 6103/6103 passed
• CHANGELOG.md updated ([Unreleased] section)
• Live API session test required (LLM serialization gate: touches message history assembly path)