v0.3.2

Added

• Intent-aware vision pipeline: the vision handoff now triages each image
  request into one of four strategies — generic, slotted, crafted, or
  decomposed — based on the adjacent user text, image count, and whether the
  image is a tool result. A new VISION_INTENT_STRATEGY config gates the
  behavior: off (generic only), slotted (force slot strategy), crafted
  (force crafted questions for single-image), or auto (default — triage
  decides per-request). The triage function is pure and deterministic so the
  chosen strategy seeds the cache key without fragmentation.
• Multi-image decomposition (DecoVQA+): when a multi-image request contains
  explicit image references (e.g. "compare the first and second image"), a
  cheap LLM call splits the user question into N per-image sub-questions, each
  neutrally phrased to defend against Visual Sycophancy. Gated by
  VISION_DECOMPOSITION_ENABLED (default true) with a configurable
  VISION_DECOMPOSITION_TIMEOUT_MS (default 3000ms). Results are cached
  in-memory per batch key so the same batch never pays twice. Failure is
  always safe — any error falls back to the slotted strategy.
• Crafted question strategy (Strategy D): for single-image complex questions,
  an LLM call reformulates the user's question into a focused, neutrally-phrased
  image-description request. The vision model never sees raw user text, which
  defends against Visual Sycophancy. Gated by the crafted triage strategy
  with a configurable VISION_CRAFTING_TIMEOUT_MS (default 3000ms). Crafting
  results are cached in-memory per input key. Failure falls back to slotted.
• Vision context extraction: ImagePart now carries adjacentText,
  isToolResult, positionInBatch, batchSize, and originalSystemPrompt.
  These fields feed the triage function and the crafted/decomposition prompts.
  New config fields control extraction bounds: VISION_ADJACENT_TEXT_MAX_CHARS
  (default 500), VISION_RECENT_MESSAGES_COUNT (default 6), and
  VISION_SYSTEM_PROMPT_MAX_CHARS (default 1000). All hot-reloadable.

Changed

• New DB index on vision_descriptions(image_hash): speeds up persistent
  cache lookups for the two-tier cache hit path. Auto-applied on next startup
  via the existing migration.

Fixed

• CI test flakiness on v0.3.1: two timing-sensitive vision tests
  (vision-handoff-background-signal and vision-handoff-integration) used
  fixed sleeps that could exceed the wait window under CI runner load. Both
  now poll for the actual condition with a 5s deadline instead of a single
  fixed sleep, matching the existing waitForModelsRequest pattern.