v0.1.2 — agent-UX patch

Patch release that doesn't change the public API or spec — just rewrites the strings agents actually read and enriches the path-rejection error envelope. Triggered by real agent feedback (Claude Code
session against an Elasticsearch cluster reflexively passing include_body: true, GPT-5.4 session hitting path-rejection confusion across npx/Docker/remote MCP).

Changed

• Tool descriptions rewritten. http_request now leads with the recommended default flow (http_request → schema + cache_id → http_read with a jq mask) and explicitly frames the cost of
  include_body: true (200KB JSON ≈ 12K tokens). http_request also gains a 2-example mini-cookbook (unknown-endpoint exploration + GitHub-issues sort/slice) and a chunked-multipart compatibility
  note. http_read gets a "lead with length" micro-tip; http_inspect calls out which schema_format to pick when. (#17)
• Path-rejection error envelope is richer. When ensureUnderRoot rejects a path, error.detail now carries field, value (original input), resolved (path.resolve(value) so traversal
  attempts surface), root, runtime ('docker' or 'host'), and hint — runtime-specific guidance text. Lets agents self-correct without guessing whether the fix is the path, the bind mount, or
  the request shape. (#19)

Added

• README "Run modes & file paths" section: decision table for npx / Docker / remote MCP, an explicit per-runtime matrix for where multipart.files[].path / download_to / save_to resolve, and
  a multipart compatibility note (chunked transfer encoding; modern servers fine, some legacy proxies may reject). (#18)
• README "Default flow" callout above the Tools table, mirroring the rewritten http_request description so in-tool docs and README don't drift.

Full changelog: CHANGELOG.md