stranded/* quarantine refs: 50 of 52 hold content that exists nowhere else

[kyle-sexton]

What

The stranded/* refs pushed during the 2026-08-07 worktree cleanup have never
had a recorded disposition — they were quarantine, and "when do these get
deleted" stayed open. Audited now, and the answer is the opposite of what
"quarantine" suggests.

52 refs across 6 repositories. 50 hold content that is on no base. 2 are
redundant.

Repository	contained	landed	unique
claude-code-plugins	0	0	44
standards	0	1	2
github-iac	0	0	2
ci-workflows	0	1	0
medley	0	0	1
dotfiles	0	0	1

Method, per ref: contained means the tip is an ancestor of origin/HEAD, so
every commit is already there. landed means the ref's RANGE patch-id (computed
--verbatim, so whitespace-only variants are not treated as the same change) is
in the base's per-commit id set — the shape a squash-merge leaves. unique means
neither held.

Why this matters

These refs were created as a safety net for 56 worktree removals, on the
assumption they were mostly redundant copies. They are not. For 50 of them, the
ref IS the work — deleting them destroys content that exists in no other
reachable place. The pending decision was framed as "when do we clean these up";
the honest reframing is "these are 50 pieces of unfinished work with no owner and
no ticket".

The two safe to delete today, both already on their base:

• ci-workflows — its single stranded/* ref
• standards — one of three

What to do with the other 50

Each needs one of three outcomes, and none of them is deletion-by-default:

1. Land it — the work is still wanted; it becomes a PR.
2. Abandon it deliberately — recorded as not-wanted, then the ref goes.
3. Keep it as an archive — the current state, which costs a ref and nothing
   else. Legitimate, but it should be a choice rather than a default.

Names carry their own signal. Fourteen are stranded/wip/*-paused-merge or
*-detached, which are mid-operation captures rather than authored work, and are
the most likely (2). Others name a live issue number — wip/1264-babysit-readiness-gate,
wip/587-fetch-comments-paused-merge, wip/926-wire-ecosystem-gates-execution —
and those are the ones where deletion would silently drop work someone already
filed a ticket for.

Reproducing

The audit is per-ref and read-only: resolve origin/HEAD, test
merge-base --is-ancestor, then compare the ref's range patch-id against the
base's per-commit id set. Same primitives as
plugins/source-control/scripts/landed-work.sh, applied to refs instead of
worktrees.

Found while closing out the stranded-work detection work (#1970).

[kyle-sexton]

Operator packet — stranded/* disposition (drain bc-f3123271)

52 refs across org repos; 50 unique-content, 2 redundant.

Host actions:

1. For the 2 redundant refs: delete after confirming content exists on a base.
2. For each unique ref: land / abandon / archive with recorded disposition (use landed-work.sh analysis where helpful).
3. Done when every stranded/* ref has a disposition and redundant pair is gone.

Cross-repo; not completable from this single-repo cloud session alone. No new issues.