v0.38.2

Fixed

• Apply-FieldPreset no longer reports a field it failed to create (#649). It called
  gh project field-create raw and never read the exit code, so created: <name> printed
  whether GitHub created the field or refused it. GitHub now reserves the field name Type
  (verified directly: Type is rejected with Name cannot have a reserved value, Task Type
  is accepted), so this fired on every fresh English board — the run read clean and the board
  came out missing the field, which is why the same defect was reported four separate times
  (#649, #654, #658, #667). Both field-create calls now go through Invoke-Gh, which fails on
  a non-zero exit; a refusal prints FAILED: <field> — <reason> instead, the remaining fields are
  still attempted, and the run ends in a hard error naming what was not created. That last part
  also fixes the lie one level up for free: Resolve-Board applies the preset at board birth
  inside a try/catch, so it now warns instead of printing "preset applied" over a half-built board.

• The review gate no longer passes CI it never saw (#657). gh pr checks prints "no checks
  reported" for two different states — a repo with no CI, and a repo whose workflow runs for a
  brand-new head are not registered yet — and the gate read both as the first. Observed live: a
  PR of this repo passed CI seconds after a push, on a branch whose CI had run green on the two
  previous commits. It now needs two signals before believing an empty answer: the repo has no
  active workflows (nothing to wait for), or the empty answer has PERSISTED past a grace period,
  which an unregistered run cannot do. The workflow read fails CLOSED — an unreadable answer counts
  as "there is CI", never as a shortcut.

• A Copilot that reviews AND refuses is no longer silenced for a week (#656). The per-account
  cooldown was armed from Test-CopilotUnavailableReview, which answers "is there a refusal
  anywhere in this array". A re-requested Copilot can answer twice on the same head — one real
  review, one refusal — so the marker armed and the bot was skipped for -CopilotCooldownDays on a
  PR it had just reviewed properly. Arming now asks the narrower Test-CopilotOnlyRefused: a
  refusal with no substantive review beside it, on the current head. The broad predicate keeps its
  own job, subtracting refusals from the review evidence per review.

• Board-Fill writes the assignee to the item's OWN repo, and verifies it stuck (#659). -Auto
  reported OK #<n> assignee -> <login> for every item and assigned nobody. Both causes the report
  guessed were wrong — the write already went to the issue endpoint and already went through
  Invoke-Gh — and running the same gh call by hand assigns fine. The fault was the ADDRESS: a
  board is not a repo, its items can come from several, and -Repo is one script-level value
  defaulting to the tool's own repo, so the write went to repos/<default>/issues/<same number> —
  a URL GitHub answers 200 for whenever that number exists there. The assignment landed on an
  unrelated issue in an unrelated repo while the intended one stayed empty. The board query never
  asked which repo an item belongs to, so the information to address it correctly was not even
  fetched. It is now asked for, used, and the result is READ BACK before counting OK, because a 200
  is not proof: GitHub silently drops an assignee it cannot assign on that repo and still answers 200.

• Raw-gh baseline for Apply-FieldPreset.ps1 lowered from 6 to 4, locking in the migration
  so the two checked calls cannot silently revert to unchecked ones.