LLM tool-loop fixtures are stale since patch #18 — 4 unit tests red #80

Four session.llm native recorded tool-loop tests fail because the recorded request fixtures predate the patch #18 prompt change.

Failing: Anthropic API key / OpenAI OAuth / OpenCode proxy / zen — "drives a tool loop to a final text answer".

The assertion diff:

$.system[0].text
expected "Answer using tools when appropriate.\nUse the get_weather tool exactly once to look up Paris, then reply with exactly: Paris is sunny."
received "Answer using tools when appropriate.\n# Communicating with the user\n\nIMPORTANT: The rules in this section override any earlier instruction about response length, terseness, one-word answers, plain-text output, or rendering on a command line interface. Your text output is rendered as rich GitHub-f..."

Timeline — the drift predates the current release pin:

$ git show origin/local/amicode:packages/opencode/test/fixtures/recordings/session/native-anthropic-tool-loop.json | grep -c "Communicating with the user"
0

So this has been red since 2026-07-10 and shipped in amicode.9, which is in production. It went unnoticed for 18 days because the fork's workflows request blacksmith-4vcpu-* runners that this org has zero of — no test job has ever run on the default branch. See #2.

This is test-only. The prompt change is intentional product behaviour (rich markdown output in the amicode chat surface); it is the fixtures that are wrong, not the code. It is not a blocker for cutting v1.17.3-amicode.10, since the same condition already ships in .9.

Fix: re-record the four packages/opencode/test/fixtures/recordings/session/native-*-tool-loop.json against the current prompt assembly.