bug: a WAL checkpoint is blind to same-process readers and can backfill past a live snapshot

[dpsiderius]

ADR-0047 (as amended on PR #723) now records this honestly; this is the
follow-up it names.

• src/vfs/shm.rs::active_reader_marks probes slot occupancy with a
  non-blocking F_WRLCK. POSIX locks do not conflict within one process, so a
  reader mark held by this process is invisible to it — the exact bug class
  bug: two Connections on one file in one process do not lock against each other — a write that reports success is silently discarded #706 exists to fix, left unfixed for this one lock. With one in-process
  WalReadLock held, active_reader_marks() returns empty.
• src/pager/checkpoint.rs folds an empty mark list into
  .unwrap_or(total_frames), so the bound becomes the whole WAL — a checkpoint
  can backfill past a live same-process reader's snapshot.
• Two in-process readers racing claim_wal_read_lock can land on the same slot,
  the second clobbering the first's published mark.

Reachability, measured: Pager::switch_wal_to_journal calls
checkpoint_passive in a loop and is reachable from the public
Pager::set_journal_mode; PRAGMA journal_mode=WAL then =DELETE works from
the CLI today and removes the -wal file. So the checkpoint path is live now.
What is still missing for harm is a second in-process reader, which needs the
unmerged embedding API (PR #705) — so not a live data-loss bug today, and it
stops being dormant the moment that pool lands.

Scope: route claim_wal_read_lock and active_reader_marks through the
in-process arbiter #706 added (src/vfs/inode_registry.rs) so a same-process
reader mark is visible and two readers cannot share a slot; make the
empty-mark-list fallback in checkpoint.rs fail closed rather than open. Tests:
a same-process reader must bound a same-process checkpoint; two in-process
readers must get distinct slots. While in this file, add the direct unit tests
inode_registry.rs still lacks (357 lines covered only by 4 oracle-gated
integration tests that no-op without the pinned oracle; SHARED ref-counting and
mutex-poison recovery were both verified by hand during review and neither is
regression-tested).

Complexity

Estimate: medium
Reasoning: The arbiter already exists; this extends it to one more
lock, tightens one fallback, and backfills unit tests.

Refs: 013/Req-4, #491, #412, #706, #723, ADR-0047