New-branch launches can be based on a stale cache without saying so

[blooop]

Destination

A branch dl cuts for a wf ticket — the shape wf hands it for every agent launch,
dl owner/repo@wayfinder/<repo>-<n> — is based on the remote's current default-branch
tip, or the launch says out loud that it isn't. Done when no arm of new-branch creation
can silently base the branch on a stale cache, and tests pin each arm.

Notes

Single-ticket map: known work, no fanning.

The happy path is already correct and already pinned — ensure_branch
(workspace_clone.py:504) fetches the requested ref, gets RefMissingOnRemote for a
brand-new wf branch, then fetches the default branch before basing on it, and
test_a_brand_new_branch_starts_from_the_current_default_branch
(test/integration/test_repo_manager_real.py:318) covers it. This map is about the
three degraded arms beside it, each of which logs a warning and proceeds from a cache
of unbounded age.

Constraint that shapes the fix: the fetch policy here was decided in
#144 and built in
#149/#150.
Launch-from-cache on an unreachable remote is deliberate and pinned by
test_an_unreachable_remote_launches_from_the_cache. So the answer is not "refuse to
launch" and not a broad +refs/heads/* fetch on the foreground path — it is making the
staleness legible to the caller.

Skills: wf-tdd to build (red before green), wf-review in fresh context.

[blooop]

Destination reached and released in v0.0.28.

The map's single build node was #245, merged as #247. Every arm of new-branch creation now answers with the base's freshness — ensure_branch returns a fresh base or a stale one naming the unrefreshed ref and the reason — and a cold launch turns a stale answer into one may be behind the remote line in dl's own output. Tests pin each arm: three unit tests for the three degraded arms plus the two fresh ones, and an integration test reproducing the observed field shape over real git.

The two constraints this map set are intact: nothing refuses to launch (#144's test_an_unreachable_remote_launches_from_the_cache passes unchanged), and no broad +refs/heads/* fetch was added — the happy path still costs one targeted fetch, two when the branch is brand new.

Worth recording as the actual cause of the reported staleness: the happy-path fix (#144/#149/#150, merged as #185) had been correct on main since 14 Aug but sat unreleased behind v0.0.26 for 117 commits, so wf — which launches through the released dl — kept cutting branches with the old interval-gated fetch. v0.0.27 shipped that; v0.0.28 shipped this map's legibility work.

Closing.