[GCdePaula]

Follows #301, now merged; this branch is rebased onto main.

Why

#301 turned a torn dispute boundary into a stop. That removed the last
production path into the dispute write-back's content-addressed miss, the
branch that stored a crossed boundary in full. Today every production
dispute feed lands on the window its boundary was loaded from:

• positioning crosses on clones and publishes the target boundary;
• nested levels and level-0 quartets below window height lie inside one
  window;
• the frontier serves the rest.

An instrumented run saw 467 positioned rulers and no production feed
beyond the positioned window. The write-back's only remaining callers were
tests that replay whole epochs on a fresh store.

This PR makes that contract explicit before the scoped review of
engine/{ruler,machine_stf,dispute}.rs, so the reviewer reads one feeder
contract instead of two.

Reading guide (commit by commit)

1. test: replay levels wider than a window on the scratch feeder.
   engine_level_root dispatches on the level's span. A level wider than
   one window (the echo golden epoch_root_r44_h48, the runner's root
   replayed under both tables, and the corpus mcycle cases) now replays
   on the scratch feeder. scratch_quartet_root is generalized to any
   stride for this. Golden values are unchanged.
2. test: pin a rejection fed at a crossed dispute boundary. This is
   new coverage. Echo rejects input 2. A middle-level root over window 2
   reverts to the directory the crossing published, matches the scratch
   replay, and ends idle on boundary 2's state. No test fed a rejection
   above boundary 0 before.
3. refactor: feed a dispute ruler only the window it was positioned
   at. This is the contract change. Feeder::Store keeps the
   positioned window and the directory ruler_at loaded and verified,
   and uses it as the revert checkpoint, as the Advance feeder does with
   its batch checkpoint. Feeding any other window panics. The constructor
   is renamed at_stored_boundary and is private. Two other changes:
   • snapshot_resumed_source_matches_template_replay is deleted. It
     drove the write-back directly, and what it compared stays pinned by
     restarted_source_resumes_a_half_built_level and
     dispute_source_matches_prototype_tree.
   • Tests and docs that described a write-back now describe the
     crossing. In node-architecture.md, the reason not to skip a torn
     boundary is now the single-window feeder.
4. refactor: delete the dispute write-back's boundary store. This
   removes Storage::commit_boundary_machine, stage_machine_store, and
   their four storage tests. After this commit, production never calls
   cm_store. publish_staged, checkout and commit_clone stay, since
   the runner and the crossing use them.
5. docs: say which stores the dispute machine can serve unaided. The
   frontier comment in dispute.rs still promised full-epoch replay from
   a store the runner never processed; it now says the machine serves only
   quartets that feed at most one input, and that wider levels replay on
   the scratch feeder in tests.

Net production code: -62 lines (snapshots.rs -67, machine_stf.rs +5).
Tests: -117 lines of storage unit tests, -28 in engine_machine.rs.

What to scrutinize

• The new single-window assert in Feeder::Store. If the root stride
  could ever exceed one window, or a production ruler fed past its
  positioned window, this assert is the loud revisit trigger. Before this
  PR, that case was a silent full store.
• The revert checkpoint is now the positioned directory itself, with no
  sync and no rehash. restore_rejected still asserts the reloaded root
  equals the revert root.

Gates

All run in the emulator v0.22 devshell, on v0.22 echo and yield images (archive_version 8), on the tree this branch has now (the rebase onto main left it byte-identical):

• just check: green. The node lib has 310 tests: fix(node): robustness pass from the Go node and feature/prt comparison #301's 314 minus the four deleted storage tests.
• just test-engine-machine: 22 passed.
• just test-node-harness: 28 passed.
• Corpus replay (computation_hash_corpus_dave_matches_release_manifest, release, against the v0.22.0 corpus): passed.
• just e2e-smoke (TEST_INSTANCE=18619): 0 of 5 failed: echo simple, chaos, kill_catchup_batched and stf_all; and yield stf_revert, whose dispute over the revert transition the honest node won.

The base branch was changed.