fix: the holes the first validated run found

[thedancingdeveloper]

Closes #121. Closes #122. Closes #123. Closes #124.

All four were found by running the previous uplift on Node B against NGMS and
watching what it actually did. Each one made the harness quieter than it
should have been about work it was losing.

#123 — the lease is held for the whole attempt

The lease was renewed only between stages, which made it a bound on the
longest stage rather than on how long a worker may be absent. The two stages
that matter — an agent thinking, and a full check suite — are both routinely
longer than it.

One 915s agent run against a 900s lease was enough. The lease lapsed 15s
before the agent returned, a second worker's claim scan retired the item
underneath it, and because every later write is owner-scoped to an owner the
row no longer had, the 15 minutes of work, the reason the item ended and its
13G worktree all disappeared without an event.

LeaseHeartbeat now beats on a daemon thread for the duration of an attempt,
in both executors. Daemon on purpose: a worker that dies must take its
heartbeat with it, because that is what makes the lease expire and the item
recoverable. release() finding itself owner-less now logs the discarded
result rather than returning silently.

The regression test is end-to-end and fails without the fix — an agent slower
than the lease, and a competing claim while it runs.

#122 — latest comes from the store the fleet writes to

Nothing calls ingest() under serve, so the EventStore the item
projections read is empty in every supervised deployment and latest was
null forever. It now prefers the audit store, grouped in SQL rather than
scanned: a "last N events" scan drops precisely the long-quiet item whose
status someone is trying to read.

#124 — a dropped pr_url is recoverable

Reconciliation was driven by the URLs the queue had already recorded, so a
pull request whose URL was lost was invisible to the one thing meant to find
it. Matching on the head branch recovers it. Bounded by construction: only a
branch an item itself recorded can match, so a human's PR cannot be adopted.

#121 — base checks do not block the request

Running the suite inside the request repeated the mistake stop had just been
fixed for. It 502'd at the proxy's 15s timeout while the build carried on with
nowhere to report, and the obvious response — retry — started a second
concurrent build. Two 502s left 26G building invisibly.

POST /api/projects/{id}/preflight/base starts a run and returns; a second
call joins the one in flight; GET polls it. check_base=true now reports
the latest run and never starts one, so a readiness read stays a read.

Gates

548 tests, ruff, mypy . — all green.