feat: refine upgrade mechanism; minor tweaks; docs - #24

Make the maintainability gate real back-pressure: raise complexity
--threshold from 1 (a no-op) to 27, the true floor. Split the files that
set it — create-task, findings-review, settings, and the board
commands/store — into cohesive modules by moving nested closures to
module scope with explicit context and extracting presentational views.
Behavior is unchanged; existing tests pass.

Restore format auto-fix in the fast loop (verify:format / verify:format:fix
via oxfmt), and expand oxlint coverage to test, scripts, and docs with
scoped overrides (no-console off for CLI scripts, no-non-null-assertion
off for tests).

Use --pushback rather than --block-new-comments for the comments gate:
every .tsx file must open with the @jsxImportSource pragma the host's
Solid transform requires, and that pragma is not a recognized machine
directive, so --block-new-comments would reject every new .tsx file on
the commit that introduces it.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…rd permission queue

Read-only helpers (intake/validator) could spawn subagents, which run as fresh
sessions that don't inherit the parent's read-only auto-approval. A nested
subagent's out-of-worktree read then blocked on a permission that never bubbled
to the board, wedging the task in "reviewing…" forever.

- Disable the subagent (task) tool for intake and validator so they stay
  single-session and read-only; they retain full direct read access.
- Resolve a permission's owning board task by walking the full parent chain, so
  worker subagents at any depth surface on the board.
- Replace the single awaitingInput marker with an awaitingPermissions list;
  add/remove entries on permission.updated/replied via a read-modify-write
  metadata mutator. Read-only helper asks are kept out of the queue.
- Board: press `p` to open the waiting-permission queue (list + jump to session),
  with toasts, a `needs you` count badge, and a footer hint.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…tall

Discovery and installation are now separate and approval-gated. At launch
Kagan only checks npm's stable latest dist-tag for global npm installs and
records an available version; nothing downloads or changes config. The
kagan.update command (board `u`, palette, /kagan-update) confirms the target,
stages the exact release via api.plugins.add (host owns the compat check),
then runs the current OpenCode executable's `plugin <exact> --global --force`
with no shell. Success requires a restart to load the new version.

Removes all OpenCode private-cache coupling: wrapper path validation,
promotion, backups, markers, disposal-time promotion, interrupted-update
recovery, and manifest/engine-range classification, plus their tests. No
migration of legacy updater artifacts (early alpha).

engines.opencode drops the moving upper bound for a minimum-only floor
(>=1.17.13); a pins guard rejects reintroducing an upper bound. Bumps
@opencode-ai/plugin and @opencode-ai/sdk to 1.17.20.

Docs, README, troubleshooting, AGENTS.md, CONTRIBUTING, and the authoritative
supervision-board Requirement 18 + design updated for global-only install,
local-alpha migration, confirmation, exact staging, restart, and rollback.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- serialize bulk task creation per project so overlapping runs can't mint
  duplicate task numbers
- roll back the worktree and branch when board-task setup or session
  creation fails, leaving no orphans
- hold the update guard across confirm+install so a second update request
  can't race a concurrent global install
- verify read-only helper ownership (parent back-pointer + owning board
  task) before auto-allowing permissions, closing a forged-role bypass

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>