v0.7.0

Board lanes you can color, limit, and edit in place, opt-in Obsidian auto-start, and a large batch
of multi-agent orchestration reliability hardening.

Added

• Board column customization. Give each lane an accent color and a soft WIP limit (with an
  unobtrusive over-limit indicator), apply a column preset at project creation (Agent Dev / Research /
  Ops / Simple) with a reset-to-preset action, and edit the board in place — rename, add, or remove
  columns directly on the board header with a live preview and role-coupling warnings. New columns are
  inserted before the terminal lane.
• Opt-in Obsidian auto-start. A new per-project setting (obsidian.autoStart, default off) that
  self-heals the vault tooling: when a skill needs the obsidian CLI and Obsidian isn't running, Loom
  launches it and waits until it's ready, then proceeds — falling back to direct filesystem access when
  it's disabled, headless, or not installed (never a hard error). Cross-platform, with an optional
  human-only launch-path override.
• Repository-path editing. Change a project's bound git repository from Settings (validated as a real
  repo; refused while a worktree session is live so an in-flight worker can't be rebound out from under).
• Spawn a worker by agent name. worker_spawn now accepts an agent's name or slug (not only its
  id), and a mistyped value returns a "did you mean …?" suggestion instead of a bare failure.
• Editable agent prompts on the Platform page and a agent_update patch surface.

Changed

• Manager sessions know where things live. An orchestrator session now starts with its project's
  absolute repo and vault paths, so it reads its notes by path instead of a slow filesystem search.
• Upward reports carry a delivery status. A worker's report (and a platform escalation) now reports
  whether it was delivered live, queued, durably boarded, or dropped — and a report wakes a parked
  manager, so completed work is never left sitting unnoticed.
• Bounded cross-project listings. The platform/audit agent and session listings are capped to fit the
  context budget, so a large workspace can't overflow an operator.

Fixed

• worker_spawn validates its inputs up front — a malformed or stale task id is rejected before any
  worktree or session is created, instead of binding a worker to a bogus task.
• worker_merge won't silently pass an empty merge. When a worker reported changes but its branch has
  nothing to merge (work was committed to the wrong place), the gate now hard-flags it for recovery
  instead of quietly marking the task done. The worker doctrine also now says, explicitly, never commit to
  main.
• Usage-limit handling on spawn is self-healing. A worker_spawn blocked by a usage limit now returns
  a retry-after deadline and the manager is auto-woken when the limit clears — no manual "retry" pokes.
• Cleaner restart resumes. After a daemon restart a session resumes as one coherent turn (the bare
  "Continue" no-op the engine emits is absorbed) and is told its file-read tracking was reset, so it
  re-reads before editing.
• Vault-tooling correctness across the board — unified board column coloring with AA-contrast-safe
  labels, Settings now preserves per-column accent/WIP on save, rate-limit holds clear cleanly across
  parked sessions, and several boot-migration and merge-recovery edge cases were hardened.