fix(abi): prune the objects recorded broken, not every risky package (#211)

## Why

A dry run of the prune on `ubuntu-2404-amd64` (pipeline 12494):

```
Dry run: 3013 object(s) left in place. Re-run with --apply to delete.
```

3013 objects on one slot. The recorded verdicts say roughly **230** of that slot's objects actually fail `dyn.load()` anywhere.

The gap is the selector: the prune used `abi_cache_risky_packages()`, the source classifier, which calls **93%** of compiled packages risky where the binaries' own symbols say **4.3%**. Applying it would have deleted ~2800 working binaries and turned each into a source fallback for every client that cannot be routed by R minor.

This is the same mistake #129 fixed in the index writer. The prune was written before the per-object verdicts existed (#200, before #128/#129) and was never brought forward.

## What changed

- Selection comes from `abi_flat_safety`, every `NOT safe` row for the slot regardless of which minor it fails under, since the flat object is served to clients on all of them.
- Matching is by **object**, `<package>_<version>`, not by package name. A package can have a safe flat binary at one version and an unsafe one at another; the old code deleted both.
- The explicit `[pkg ...]` override is unchanged in behaviour and still matches by package name, which is what you want when naming a package by hand.
- The refusal guard moves with the selector. An empty result now means the slot's flat-safety backfill never ran — the exact state rhel8 was found in today — and pruning nothing while reporting success would have hidden it.

## Not changed

Still a dry run unless `PRUNE_APPLY=yes`, and still refuses to delete an object whose package CRAN no longer ships, since there would be no source to fall back to.

## Ordering note

A prune converts each deleted binary into a source fallback. bincraft #131 fixes source fallbacks of *archived* versions being advertised as binaries, which breaks `uvr` outright. That should land before any large prune, or this will manufacture new instances of it.

Reviewed-on: #211

.crow/prune-stale-flat-binaries.yaml

@@ -1,11 +1,18 @@
-# Remove flat-slot binaries belonging to ABI-risky packages, then re-index.
+# Remove flat-slot binaries recorded unable to load under some R minor, then
+# re-index.
 #
-# `upload_single_binary()` writes an ABI-risky package only to the per-minor
-# slot, so a risky package should have no flat binary at all. Packages built
-# before they were classified risky still do, and the flat slot is what every
-# client that cannot be routed by R minor receives. That binary was built under
-# whichever R was primary at the time, so it dies at load under any minor that
-# dropped a symbol it uses:
+# The list comes from `abi_flat_safety`, so it names the objects the binaries
+# themselves report as broken, not the ones a source heuristic suspects. That
+# distinction is the whole cost of the operation: the heuristic calls 93% of
+# compiled packages risky and would delete roughly 3000 objects per slot to fix
+# about 230.
+#
+# `upload_single_binary()` has written a non-portable package only to the
+# per-minor slot since v5.3.0, so this is a one-time cleanup of what was
+# uploaded before that. The flat slot is what every client that cannot be
+# routed by R minor receives, and such a binary was built under whichever R was
+# primary at the time, so it dies at load under any minor that dropped a symbol
+# it uses:
 #
 #   base64enc.so: undefined symbol: SETLENGTH
 #