file_scan: refuse unsupported-fs roots up front instead of failing silently

[martinus]

Pre-launch robustness: make the "pointed oans at the wrong filesystem" case
honest, since it's a predictable first-touch mistake for new users.

The problem (observed)

Point oans at a directory on ext4/tmpfs and, on a modern kernel where
FS_IOC_GETFSUUID succeeds, it would:

1. print a single Warn: filesystem ... is not known to support deduplication
   (easily lost in progress output),
2. walk the tree anyway, hashing every file,
3. fail every extent op with per-file fiemap: Operation not supported spam,
4. and still exit 0 with Nothing to deduplicate — looking like success.

(The other case — where get_uuid outright fails, e.g. older kernels — was
already handled loudly by filescan_seed_failed(); this fixes the modern
"succeeds but unsupported" gap.)

The fix

In check_file(), when the identified fs isn't btrfs/XFS, refuse the seed root
via seed_reject() with a clear Skipping X: its filesystem is not btrfs or XFS message instead of returning true. Consequences:

• All roots unsupported → the existing filescan_seed_failed() path fires:
  loud error, exit 1, and no FIEMAP spam (the tree is never walked).
• Mixed → supported roots scan normally; unsupported ones are skipped with
  the message.
• locked_fs is now committed only after the support check (the UUID lands
  in a local first), so a rejected root can't pollute the fs-lock for a later
  supported root.

Behaviour change to note

A filesystem that reports a UUID but isn't btrfs or XFS is now skipped
rather than attempted. oans only deduplicates btrfs/XFS (via FIDEDUPERANGE),
so this converts a misleading exit-0 "success" into an honest, actionable
failure. If we ever want to support another dedupe-capable fs, add its magic to
is_fs_supported().

Before / after (oans -dr on a tmpfs dir)

Before: Warn: + fiemap: Operation not supported ×N + Nothing to deduplicate, exit 0.
After: Skipping <dir>: its filesystem is not btrfs or XFS ... + Error: could not determine the filesystem for any given path ..., exit 1, no FIEMAP spam.

Testing

• New tests/integration/test_unsupported_fs.py (skips when the system temp
  dir happens to be a reflink fs).
• scripts/verify.sh green: clean build, 91 tests, valgrind scan+dedupe+replay
  smoke clean.

🤖 Generated with Claude Code

[martinus]

Ran a /simplify pass and applied its findings:

• test now reuses the harness scratch_fstype() helper instead of a local duplicate (dropped the redundant subprocess import and the double probe);
• reworded the generic end-of-run error in oans.c from "could not determine the filesystem…" to "none of the given paths are on a filesystem oans can deduplicate…", since with this change we now often have determined the fs — it's just unsupported.

The core check_file change was confirmed at the right depth (reuses seed_reject/filescan_seed_failed, commits locked_fs only after the support check). verify.sh still green (91 tests + valgrind).