- Remove unused AggregateChannel compat wrapper from test files; switch
  all DeltaChannel test aliases to import from channels.delta directly
- Drop dict-reducer tests (tested AggregateChannel type-inference, not
  relevant to this DeltaChannel-only branch)
- Add value: Value | Any annotation to AggregateChannel.__slots__ (mypy)
- Add isinstance asserts for DeltaChannel before replay_writes calls
- Remove now-unused _math_compat import and clean up PostgresSaver import

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

Restores Postgres checkpointer support to both benchmark sections.
Uses local Postgres at port 5441. Each run gets a fresh table slice
via DELETE before and after to avoid cross-contamination.

Key Postgres results at 500 turns:
  freq=1   → 8.3ms reads  (full snapshot every write)
  freq=5   → 4.8ms reads  (bounded replay, fewer large blobs to fetch)
  freq=10  → 5.3ms reads
  freq=50  → 6.7ms reads
  freq=inf → 123.7ms reads (full ancestry walk)

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…t ext type

Replaces the write-count (_write_count) snapshot mechanism with a clean
step-based approach: pregel's create_checkpoint fires snapshots every N
pregel steps regardless of whether the channel was written (eager).

Key changes:
- snapshot_frequency=None (default) for pure delta; int N for snapshot every N steps
- DeltaChannel.checkpoint() always returns DELTA_SENTINEL; snapshot logic
  lives in create_checkpoint which has the step number
- create_checkpoint bumps channel version via get_next_version when the
  channel wasn't written at a snapshot step (eager: always stores the blob)
- _DeltaSnapshot NamedTuple registered as EXT_DELTA_SNAPSHOT (code 7) in
  the msgpack serde — no dict key collision, type tag does the dispatch
- InMemorySaver and PostgresSaver _get_channel_writes_history updated:
  _DeltaSnapshot blobs collect pending_writes before terminating (they
  encode the NEXT step's transition, not subsumed by the snapshot unlike
  pre-delta migration blobs)

Tests confirm:
- Snapshots fire at every N steps even when channel has no write that step
- Correct accumulated state after reconstruction from snapshot + replay

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…annel reconstruction

InMemorySaver.prune now walks the parent chain from the latest checkpoint
and keeps all ancestors whose writes are still needed for DeltaChannel
reconstruction — i.e. ancestors with DELTA_SENTINEL blobs where no
_DeltaSnapshot has been written yet in the kept ancestry.

Once a _DeltaSnapshot blob exists for all sentinel channels (provided by
snapshot_frequency), the walk terminates and older checkpoints plus their
writes and blobs are safely deleted.

With snapshot_frequency=None (pure delta), all intermediate checkpoints
are retained since the full write history is required. Setting
snapshot_frequency=N bounds the kept ancestry to N steps.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…as DELTA_SENTINEL

In async (default) durability mode, put_writes is fire-and-forget. If
checkpoint_writes fails but put succeeds, the sentinel blob has no
backing writes — reads silently reconstruct empty/wrong state.

Fix: track futures returned by put_writes submissions in
_pending_write_futs. Before committing a checkpoint that contains any
DELTA_SENTINEL blob, call .result() on all pending write futures,
blocking until they complete. This ensures checkpoint_writes are durable
before the sentinel blob is committed.

The flush only triggers when DELTA_SENTINEL is present, so graphs
without DeltaChannel channels are unaffected. Snapshot steps
(_DeltaSnapshot blobs) are self-contained and do not need the flush.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…aSnapshot correctly

The base saver's _get_channel_writes_history walk terminated at any
non-DELTA_SENTINEL blob without collecting that ancestor's pending_writes.
This was correct for pre-delta migration blobs (which subsume their own
writes), but wrong for _DeltaSnapshot blobs: the snapshot captures state
AT the ancestor, while pending_writes encode the NEXT step's transition
and must be collected before terminating.

Applies the same fix as was already applied to InMemorySaver and
PostgresSaver: when the ancestor blob is a _DeltaSnapshot, collect its
pending_writes first, then terminate with the snapshot as seed.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>