Per-provider vision blocks; parser repairs local models' almost-JSON

config.toml now carries a [vision.<provider>] block per backend —
model/base_url/key_env — with vision_provider picking the active one,
so the committed file documents every recipe and switching is a
one-line flip. Only the active block applies; typo'd block names and
keys warn like every other config mistake.

First real Ollama run (qwen2.5vl:7b) surfaced what local models emit:
almost-JSON with trailing commas. parse_vision_response now makes one
cheap repair pass before declaring a response unusable.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g

config.toml

@@ -3,12 +3,19 @@
 
 photos_dir = "photos"
 data_dir = "data"
-model = "claude-sonnet-5"
 rate_limit_seconds = 2.0
 
-# Vision backend. The default ("anthropic") reads ANTHROPIC_API_KEY.
-# "openai-compatible" covers OpenAI, OpenRouter, and local runtimes:
-#   vision_provider = "openai-compatible"
-#   vision_base_url = "http://localhost:11434/v1"   # e.g. Ollama
-#   model = "qwen2.5vl"                             # a vision-capable model
-#   vision_key_env = ""                             # local: no key needed
+# Which vision backend reads your shelf photos. Both recipes below stay
+# on file; this line picks one.
+vision_provider = "anthropic"
+
+[vision.anthropic]
+# reads ANTHROPIC_API_KEY from the environment
+model = "claude-sonnet-5"
+
+[vision."openai-compatible"]
+# OpenAI, OpenRouter, or a local runtime (Ollama, LM Studio, vLLM).
+# key_env names the env var holding the key; "" = endpoint needs none.
+base_url = "http://localhost:11434/v1"
+model = "qwen2.5vl:7b"
+key_env = ""