Background

#29 introduced two families of local recovery refs written before an auto-rollback's git reset --hard:

1. refs/heads/attempt-preserve/<run_id>-<sha> branches parking committed work (preserve_commits), and
2. refs/attempt-preserve-dirty/<run_id>-<baseline>-<attempt> refs parking the uncommitted working-tree snapshot (snapshot_worktree via _preserve_attempt_worktree).

#32 added bounded retention for family 1 (ScmPolicy.preserve_keep, pruned at run start) and explicitly fenced family 2 out of its scope.

Problem

Nothing ever removes the refs/attempt-preserve-dirty/* refs. Every dirty auto-rollback attempt creates one (keyed by attempt number, so retries each add a new ref), and they accumulate unbounded exactly like the branches did.

Scope / severity: low. Same as #32: these refs are local-only (never pushed) and, living outside refs/heads/, they don't even clutter git branch output — so the nuisance is smaller. But they still add ref-walking overhead forever and hold snapshot commits alive against git gc indefinitely.

Proposed resolution

Extend the run-start bounded retention to this family:

• Reuse the existing scm.preserve_keep budget (simplest — one knob governs both recovery-ref families), or add a sibling field if independent tuning is wanted.
• These are not branches, so delete via git update-ref -d (or batched git update-ref --stdin), not git branch -D.
• for-each-ref --sort=-committerdate refs/attempt-preserve-dirty/ works unchanged: each ref points at the snapshot commit created at preservation time, so committer date is the park time here.
• Keep the Cap the lifecycle of attempt-preserve/* recovery refs (follow-up to #29) #32 hardening: per-ref best-effort deletes (one stuck ref must not wedge the tail), full-refname handling, prune failures journalled and never blocking run start.

Acceptance criteria

• refs/attempt-preserve-dirty/* refs are pruned to a bounded count at run start, keeping the most recently created.
• Retention is configurable; 0/disabled keeps every ref (consistent with preserve_keep semantics).
• Pruning only ever touches refs/attempt-preserve-dirty/* — never refs/heads/* (the Cap the lifecycle of attempt-preserve/* recovery refs (follow-up to #29) #32 prune already owns refs/heads/attempt-preserve/*).
• Tests cover: keep-N boundary, ordering, disabled = no-op, a stuck ref not wedging the rest, and branches being untouched.

Notes

Non-blocking follow-up to #32; surfaced during its review as the symmetric gap for the second ref family.