fix(master): compact and version-prune the store registries alongside _stats

[beinan]

Problem

_registry.rollout.lance (and _registry.generic.lance) is the shared directory of stores every worker consults. Workers write it on every store create/delete as a delete plus an append (upsert is delete-then-append so a retried create is idempotent). Nothing ever compacts or version-prunes it.

Lance keeps every manifest until cleaned, and each manifest lists every fragment, so the table grows quadratically. Production today: version 34,504, ~840 KB per manifest, ~29 GB of manifests for a 10k-row three-column table, with one fragment per row.

Every POST /rollouts does contains() (a checkout_latest that reads that head manifest) then upsert() (two commits, each writing a new 840 KB manifest); every GET /rollouts/{name} does another checkout_latest. Measured on prod: creates average 47 s (201), 83 s when the client has already timed out and retried (409), and fail with 500 when ADLS throttles mid-way. GET /rollouts/{name} averages 0.7 s for a row lookup.

Fix

RolloutRegistry gets the same compact() / detached RegistryCleaner::cleanup() / reload() split as StatsStore (#272), plus maintain() for callers without a shared lock.

The master runs it for both registries in the scanner's existing maintenance round (every STATS_MAINTENANCE_EVERY_N_SCANS scans, under the stats-writer lock so only one master rewrites). Compaction is bounded by MAINTENANCE_TIMEOUT; the cleanup runs with the lock released. Workers keep appending throughout: Lance's commit retry carries their appends past the rewrite, and the grace window (STATS_HISTORY_TTL_SECS) keeps any version a worker may still hold open.

The first pass on prod will delete ~34k manifests; with #272 that no longer blocks anything.

Verification

• maintain_bounds_versions_and_preserves_rows: 20 upserts (≥40 versions) → compaction folds fragments, cleanup removes versions, all 20 rows and further writes work.
• cleanup_runs_detached_from_the_registry: writes land while the cleaner is outstanding and survive.
• maintenance_does_not_lose_concurrent_worker_upserts: a second handle on the same URI (the worker) creates stores before, between and after the master's compact/cleanup; both handles see all 26 rows.
• etcd-backed master suite --include-ignored: 72 passed. Clippy -D warnings, fmt clean.

🤖 Generated with Claude Code