T6: exactly-once acceptance tests + smoke for incremental sink reads

[philcunliffe]

Implements task T6 of the incremental-sink-reads change set (LLP 0042).

The exactly-once proof. T4/T5 ship comprehensive unit suites, but those stub
storage and so cannot prove the design's load-bearing claim: that a row-resident
_hyp_ingest_seq watermark survives the two cache rewrites that motivate the
whole design (LLP 0039) — a retention front-prune and a compaction generation
swap.

What this adds

• test/core/sink-incremental-acceptance.test.js — a deterministic acceptance
  suite driving the real kernel cache (createQueryStorageService), the
  real retention enforcer, real maintainCache compaction, both
  sinks (central forward + core local-fs blob via their own
  activate/create), and the real sink driver's outbox respool. Eight tests:
  • forward + blob: ~0 bytes on a no-new-rows tick, ~N on an N-new tick;
  • forward + blob: exactly-once across a retention front-prune;
  • forward + blob: exactly-once across a compaction generation swap
    (asserts the generation directory actually swapped);
  • forward: watermark composes with the driver-outbox respool — the replay
    reads only the un-acked suffix with a stable X-Hyp-Batch-Id;
  • blob: a lost watermark after a durable PUT re-PUTs the same object key
    (idempotent overwrite, no duplicate blob).
• hypaware-core/smoke/flows/incremental_sink_compaction.js — a hermetic
  acceptance smoke proving the blob sink (parquet + local-fs) reads straight
  through a compaction generation swap via the real driver across three ticks.
• Adds an explicit {#exactly-once-argument} anchor to LLP 0040 §5 so the new
  @refs resolve.

Checks

npm test (1446 pass / 1 pre-existing skip), npm run typecheck, npm run lint
all green; npm run smoke -- incremental_sink_compaction passes.

Task-Id: T6

🤖 Generated with Claude Code