fix(sweep): merge grounding onto survivor on dedup collapse (#56)

[jimador]

Closes #56.

The dream-loop dedup sweep collapsed near-duplicate propositions by flipping the
losers to STALE with a pure status change. STALE is excluded from retrieval, so
each loser's grounding and provenance became invisible the moment it was retired
— evidence silently lost on every collapse. The ingestion-time reviser already
folds grounding onto the survivor; the sweep was the gap.

What changed

Fold the loser's evidence onto the survivor before retiring it. Policy decides
intent, the runner does the IO (it is the only layer with a repository handle),
keeping the existing SPI split:

• SweepAction.MergeInto(survivorId, thenStatus) — new collapse outcome.
• MergingSweepPolicy — reads MarkReason.Duplicate.survivorId, returns
  MergeInto; falls back to a plain STALE transition when pinned, unmarked, or
  when no usable survivor id is present. StatusTransitionSweepPolicy is
  untouched.
• DefaultCollectorRunner handles MergeInto: loads the survivor, folds the
  loser's grounding + provenance + sourceIds onto it (append + distinct via the
  existing Proposition helpers), bumps reinforceCount by 1 to match the
  reviser's per-merge convention, saves the survivor, then retires the loser and
  emits the same status-changed event the plain path emits. A missing survivor
  falls back to a plain retirement (logged at WARN, since the audit row looks
  like an ordinary transition). Dry runs mutate nothing.
• Opt in via CollectorRunner.Builder.withDuplicateDetection(), which pairs a
  DuplicateCollectorStrategy with MergingSweepPolicy. It defers to a policy
  set explicitly via withPolicy(), and the global builder default stays
  StatusTransitionSweepPolicy, so the decay/stale path is unchanged.

Review notes

Adversarial review raised the production save() path: findOrPersist dedups
by (contextId, text), so an in-place update could in theory be hijacked by a
same-text sibling. It's unreachable in practice — save() dedups identical text
at creation, so two same-text nodes can't coexist, and near-duplicate collapse
operates on different-text propositions whose exact text never collides. No
hot-path guard was added.

Caveat for adopters: save()'s dedup contract assumes a (contextId, text)
uniqueness constraint as a cross-instance backstop (declared in the test
SchemaCatalog). A deployment without it loses that backstop under concurrent
writers; worth applying the constraint in production schema.

Tests

• MergingSweepPolicyTest — decide() emits MergeInto for a Duplicate mark;
  falls back to STALE when pinned/unmarked/blank-or-self survivor.
• DefaultCollectorRunnerTest — live merge folds grounding/provenance/sourceIds
  onto the survivor (deduped) with reinforceCount + 1, then retires the loser;
  dry run mutates nothing; missing survivor falls back cleanly; multi-loser
  cluster absorbs every loser's evidence; withDuplicateDetection defers to an
  explicit policy.
• CollectorDedupMergesGroundingTest (integration) — collapses a real cluster
  and asserts the survivor absorbed the loser's evidence while the loser is STALE.