test: add live end-to-end validation harness

.gitignore

@@ -66,3 +66,10 @@ __pycache__/
 *-config-with-keys*
 *-with-keys*
 .rocketride/
+
+# E2E test outputs
+demos/e2e-tests/e2e-results.txt
+demos/e2e-tests/e2e.csv
+!demos/e2e-tests/config-ollama.txt
+demos/e2e-tests/config-groq.txt
+demos/e2e-tests/config-anthropic.txt

demos/e2e-tests/config-anthropic.txt.example

@@ -0,0 +1,19 @@
+# E2E validation — Anthropic (PR #71 / issue #63)
+#
+# Get a key at https://console.anthropic.com/settings/keys (needs billing).
+#
+# What this config is for: verifying the adaptive-thinking request shape.
+# The review found the new code path is unreachable for every model bundled
+# in models.yaml — all 7 are 4.5-or-earlier and take the extended-thinking
+# branch. To exercise the adaptive branch you must set a 5-series model
+# explicitly, which llm:set-model permits with a warning.
+#
+# Swap `model` below to claude-opus-4-5-20251101 to test the EXTENDED path,
+# or to claude-sonnet-5 / claude-opus-5 to test the ADAPTIVE path.
+
+provider=anthropic
+anthropic_api_key=sk-ant-REPLACE_WITH_YOUR_KEY
+model=claude-haiku-4-5-20251001
+temperature=0.0
+max_tokens=300
+timeout_seconds=60

demos/e2e-tests/config-groq.txt.example

@@ -0,0 +1,17 @@
+# E2E validation — Groq (PR #73 / issue #51)
+#
+# Get a free key in ~2 minutes at https://console.groq.com/keys
+# Free tier, no credit card. Replace the placeholder below.
+#
+# What this config is for: the code review found that a plain llm:chat never
+# sends reasoning_format, so Groq falls back to its documented default "raw"
+# and emits thinking inside <think> tags in the visible content. Nothing
+# strips them. This config is what proves or disproves that against the real
+# API — it cannot be settled without a live call.
+
+provider=groq
+groq_api_key=gsk_REPLACE_WITH_YOUR_KEY
+model=openai/gpt-oss-20b
+temperature=0.0
+max_tokens=300
+timeout_seconds=60

demos/e2e-tests/config-ollama.txt

@@ -0,0 +1,8 @@
+# E2E validation — local Ollama. No API key required.
+# Verifies the full stack end to end without touching a paid provider.
+provider=ollama
+model=qwen2.5vl:3b
+base_url=http://localhost:11434
+temperature=0.0
+max_tokens=200
+timeout_seconds=120