feat: parallel v2 phase 2 — Ordered() via batch-aligned re-sequencing

[tr1v3r]

Summary

Phase 2 of docs/proposals/parallel-v2.md: Ordered() — opt-in order preservation for parallel sections.

How it works

• Parallel(n).Ordered()... — feeder stamps input indices, workers forward Filter-dropped elements as holes (batch boundaries/lengths stay intact), consumer re-sequenences batch-keyed: an aligned batch yields straight through with zero per-element bookkeeping; holes advance the expected index.
• Ordered() without a parallel section is a no-op; Parallel(n) resets to unordered.

Design journey (documented in the commit)

Three candidate re-sequencers were built and measured:

1. next-index map — deadlocks on Filter holes (caught by the A3 property test)
2. smallest-pending — emits early under batch arrival races (caught by property test)
3. ring buffer — corrupts when in-flight distance exceeds window (consumer-held partial batches extend it; caught by a sum-mismatch stress)
4. batch-keyed pending + hole placeholders — correct (30-round adversarial stress) and fast

Acceptance (Apple M3 Pro)

gate	target	measured
A3 order property	ordered == serial, element-for-element	✅ 30 random pipelines + adversarial batch-head filtering + all-filtered
ordered overhead (near-free, gate ≤3×)	≤3× serial	2.0× (naive map was 8×)
A2 heavy speedup	≥3×	✅ 3.6× unchanged
short-circuit/cancel	no hang, no leak	✅ incl. 100k-element Limit(5)

Verification

• go test ./... -race green; golangci-lint 0 issues
• Coverage ~100% (one benign timing branch in the ordered closer occasionally reads uncovered)