bug: VertexSynthesizer HTTP timeout (120s) too short for large curation prompts

[mpeter]

Summary

dewey curate fails with context deadline exceeded (Client.Timeout exceeded while awaiting headers) for vaults with rich content. The VertexSynthesizer HTTP client is configured with a 120-second timeout, but large curation extraction prompts (30K-40K input tokens) consistently take 120-260 seconds to complete on Vertex AI.

Reproduction

Configure a knowledge store with 9-10 documents containing rich markdown (dossiers, agreements — ~34K input tokens). Run dewey curate. Observe:

❌ <store>: LLM extraction failed: vertex AI request: Post "https://...": 
context deadline exceeded (Client.Timeout exceeded while awaiting headers)

Timing: the failure occurs at exactly 120 seconds, confirming the HTTP client timeout is the cause.

Root Cause

llm/vertex.go:

client: &http.Client{
    Timeout: 120 * time.Second, // Match OllamaSynthesizer timeout.
},

The Ollama timeout is appropriate for local inference. Vertex AI with Claude Sonnet processing 34K input tokens takes 120-260 seconds — regularly exceeding the cap.

Fix

Increase the Vertex HTTP client timeout to 300 seconds. Vertex API calls have different latency characteristics than local Ollama inference and warrant a longer timeout.

Timeout: 300 * time.Second,

Context

Discovered on Vertex AI with claude-sonnet-4-6@default in us-east5. A direct curl to the same endpoint with the same payload completes in 85-261 seconds depending on load. The 120s Dewey timeout consistently cuts off large extractions.