feat(model-host): sbproxy models prune to reclaim unreferenced weight blobs

[rickcrawford]

What

sbproxy models prune — reclaim content-addressed weight blobs that no cached artifact references (orphans from an interrupted pull, or shards a removed model no longer keeps alive).

• Runs under the same collection lock as the cache-budget sweep, so it never races a concurrent pull's promote.
• Reclaims exactly the unreferenced blobs; a referenced (possibly shared) blob survives while its last reference remains.
• --dry-run reports reclaimable bytes + orphan count and deletes nothing.
• Cold / already-clean cache is a no-op.
• --cache-dir overrides the target, else the configured cache dir (with -f) then the platform default — same resolution models list uses.
• JSON (models.prune, schema_version: 1) and text output.

Live-validated: cold cache → 0 B; planted orphan → dry-run reports 4096 B / 1 blob and leaves it; real prune reclaims it; a referenced blob survives.

Context — what WOR-1862 actually needs

Investigating the ticket surfaced two things worth recording:

1. The store is already content-addressed with cross-model dedup. promote_blob stores each file at blobs/sha256/<digest> and drops a duplicate partial when the blob already exists, so two serve entries sharing a base model already store its shards once. Resumable downloads (partials/ + resume metadata) also already exist. That scope bullet is largely satisfied by existing code.
2. hf-hub 1.0 does not bundle Xet. Its only features are blocking, rustls-tls, socks; Xet lives in a separate hf-xet 1.5.3 crate with its own CAS/chunking model. So "hf-hub 1.0 with the Xet feature" (the ticket premise) is inaccurate, and swapping our working resumable-HTTP transport for hf-hub alone adds little. The real Xet migration is a larger, uncertain effort that deserves its own evaluation.

So this PR ships the clean, self-contained, offline part of 1862 (the prune command, satisfying the "prune on a cold cache is a no-op and on a dirty cache reclaims exactly the unreferenced bytes" acceptance criterion). The transport migration is left for a focused follow-up; findings are on the ticket.

Tests

Unit: cold no-op, dry-run non-destructive, exact reclamation, referenced blob survives, second prune is a no-op. Full gate green: nextest 6985/6985, clippy -D warnings, rustdoc, doctest, llms-full regen.