Fix node→pivot write loss in set/delete/set sequences

[benitogf]

Summary

Closes #65 — a node-originated write is no longer lost when a key is set, deleted from the pivot, then set again on the node.

• The node's local write now reliably advances its version vector, so the pivot accepts it instead of mistaking it for an already-seen duplicate.
• Closes two timing windows that caused the version vector to stay stale: a cross-operation bookkeeping mix-up between a concurrent pulled delete and a local set, and the bump landing after the change was broadcast (so the peer push read a stale vector).
• Makes the end-to-end sync tests synchronise deterministically on observed sync events — exact-count wait groups — instead of time-based polling or deadlines. Now that the version-vector fix removes duplicate deliveries, each operation produces an exactly predictable number of events, so the tests can't flake or hang under load (the original CI failure was a timing-related hang).

Depends on the merged github.com/benitogf/ooo operation-aware post-write hook (benitogf/ooo#139); go.mod is pinned to it.

Test plan

• Set → delete-from-pivot → set-on-node on the same key: the node's write reaches the pivot (no divergence).
• End-to-end sync tests pass deterministically with no time-based waiting (no sleeps, no polling, no deadlines).
• Full suite green under the race detector; validated ~6900 race iterations with zero divergence (was ~1/300).

---

🤖 Generated with Claude Code

[CBosch101]

Fixes the node→pivot write loss by making the local VV bump operation-aware (consume only the mark matching this write's own op) and moving it onto ooo's AfterWriteOp so it lands before the broadcast; rewrites the cluster test to converge-or-fail-fast instead of counting non-deterministic async events. The two core race-fix claims hold against the ooo source, and both product consumers build clean against this PR. Approving.

Verified

• ooo AfterWriteOp fires before the storage broadcast and before the legacy AfterWrite, and emits only "set"/"del" (ooo storage/layered.go:47-58,555,748) — substantiates both the "bump before broadcast" and "caller AfterWrite still sees post-bump VV" claims, and makes the op=="del"-else-set dispatch total.
• Op dispatch lines up with the mark setters: trackSet→bumpSkipSet, trackDelete→bumpSkipDelete (sync.go:552-565); each pulled write later consumes its own matching mark, so the op-specific consume introduces no mark leak.
• Breakage check — idnerdidx/bulk and idnerdidx/bundle both go mod tidy + go build ./... clean against this PR's pivot (which drags in the bumped ooo). Only pre-existing environmental failures remain, identical to baseline: libpcsclite/ebfe/scard (both repos) and -lstreamdock link in bundle/collector. pivot-adjacent packages test clean; bulk's lone failure (pivot/router TestRunUploadJob_PanicSurfacesAsError) fails identically at baseline.
• The PR's failing check is a pre-existing flake, not this change: macos failed on TestClockDriftScenario (not in this diff) under -race -failfast -timeout 60s; the same commit's push workflow run passed all three platforms, and it passes locally under -race -count=3.
• pivot suite green locally under -race (cluster sync ×3 + full suite). Did not reproduce the full ~6900-iteration figure — sampled only.

No blockers.