vectr 1.2.0

Proactive context injection and daemon availability under load. All new
behavior is opt-in; with no new flags or consent given, behavior is
unchanged.

Proactive context (new, opt-in)

• New vectr proxy command: a localhost, Anthropic-API-shaped proxy that
  deterministically injects relevant working-memory notes into an AI
  agent's request context when their trigger conditions match — no reliance
  on the agent voluntarily calling recall. Injection is consent-gated (a
  proactive master switch plus per-launch consent for the proxy channel)
  and fully observable: injection counts and end-to-end state are exposed
  in /v1/status and rendered in the proxy banner.
• Scored recall behind a new /v1/proactive endpoint; injections are
  budgeted, deduplicated, and fail open (an injection-path error never
  blocks the underlying request — it is logged and bypassed).
• Org-wide artifact cache and exact-match response cache wired into the
  proactive path.

Daemon availability

• /v1/index now runs off the event loop (threadpool), so a full-workspace
  index call no longer blocks every other request for its duration.
• New reindex_in_progress field in /v1/status: a lock-free, always-cheap
  signal that bulk index work (an explicit index or the watcher's coalesced
  batch) is running right now.
• Per-language index statistics no longer trigger a full metadata scan on
  every status call while the index is changing; chunk totals read the
  vector store's native count directly.

Watcher

• Live file events now honor .gitignore exactly like the bulk indexer
  (previously only .vectrignore was consulted on create/modify/delete/move
  events, so a gitignored file could enter the index through a live edit).

Fixes

• vectr forget --all sweeps the current cache layout, not only the legacy
  nested layout.
• The proxy banner's status probe retries with a longer timeout instead of
  reporting a transient failure.