v2.7.0

This release rounds out the MCP surface (the agent-facing side of journal),
makes decisions & todos first-class, and adds project tooling — a docs site,
a curated release pipeline, and the agentic-workflow layer.

Added

• MCP synth tool. The MCP server now exposes synthesis as a synth tool so
  MCP clients (e.g. an agent drafting a weekly Slack summary) can run
  weekly|daily|meetings|decisions|stale synthesis jobs without shelling out to
  the CLI. By default the tool calls the synthesis provider and returns the
  generated text without writing a file (persist: false); set persist: true to
  also write the draft note to disk, mirroring journal synth --write. Honors
  synth_provider, local_only, and voice_profile exactly as the CLI does;
  returns a clean {"error":"…"} when synthesis is unavailable.
  Optional scoping params: kind (default weekly), days, project, date.
• MCP ask tool — runs retrieve→synthesize and returns
  {answer, citations} (grounded text + path:line references) so clients get a
  direct "what did I decide about X" answer instead of raw chunks. When no chunks
  match, returns "No relevant notes found." rather than calling the model
  ungrounded. Honors local_only / provider availability with a clear error
  (#32).
• MCP stats and today tools — the same stable JSON as
  journal stats --json / journal today --json, for "how's my note volume" and
  "what does my day look like" (#31).
• MCP resources — journal://today, journal://recent, and
  journal://projects/{slug}/index expose raw Markdown as addressable context via
  resources/list / resources/read, so clients can pull journal context without
  orchestrating tool calls. URIs are stable across server runs
  (#37).
• MCP prompts — weekly reflection, decisions review, and project status
  are exposed as first-class, one-click prompts with journal context pre-assembled
  (reusing synth.AssembleWeekly / AssembleDecisions)
  (#44).
• journal tags — list distinct #tags with usage counts, and
  journal tags rename <old> <new> rewrites a tag across all notes, re-indexes the
  affected files, and auto-commits (--dry-run previews). Boundary-aware so #foo
  isn't matched inside #foobar (#33).
• First-class decisions & todos. Dedicated capture commands, crisp
  date/statement/citation rendering, proactive surfacing in journal today, and
  resolution notes on journal done. Fully backward compatible — existing
  @decision/@todo/@done blocks parse and surface unchanged
  (#47).
• Documentation site at https://journal.eamann.com — an mdBook site
  (Catppuccin Latte) with warm, second-person prose across 14 chapters (install,
  capture, search, synthesis, meetings, configuration, integrations). Additive: the
  docs/ tree is untouched (#38).
• One-click releases. A prepare-release workflow (Actions → Prepare Release →
  version) finalizes the ## [Unreleased] CHANGELOG section to the version, tags,
  and triggers the Release workflow; and the GitHub Release notes now come from the
  curated CHANGELOG section (scripts/changelog-section.sh + GoReleaser
  --release-notes) instead of an auto-generated commit list. See
  docs/RELEASING.md. Deterministic, project-specific tooling.
• Agentic-workflow layer + CLAUDE.md. GitHub Actions for agent-ready issues
  (auto-label → trigger → PR), a planning-approval gate for high-complexity work,
  @claude PR feedback (gated to write-access collaborators), and a retro flow;
  plus issue/PR templates, a label set, and a contributor guide grounded in the
  real layout. Contributor/CI tooling — no change to the shipped binary
  (#25).

Changed

• Better LLM-as-reranker quality. The rerank prompt is now a structured
  relevance rubric (0/5/10 anchors) instead of "reply with only the number", and
  score parsing is a robust multi-strategy parser (N/10 fraction → labelled
  Score: N → last in-range number, skipping negatives). More reliable precision
  on the optional reranking path (#36).

Fixed

• journal index --watch survives transient embed failures. A per-file embed
  error during a watch pass is now logged and skipped (continue) instead of
  aborting the whole batch; only context cancellation still stops the run.
  Debounce coalescing of rapid edits is now covered by tests
  (#35).
• Indexing retries transient Ollama embed-server failures (sporadic 400 /
  EOF) instead of failing the run, hardening large journal index passes
  (#7).