feat(bet1): productionize reuse-under-drift + validate on a real learned-GNN trajectory (ADR-202 WIN)

[shaal]

BET 1 productionize — fixed-topology reuse + periodic rebuild on a real learned-GNN trajectory

Closes ADR-200's named open frontier (next-step #4). Wires the reuse-under-drift policy into the production ruvector-diskann loop behind a feature flag, and validates it on a genuine learned-GNN embedding trajectory (contrastive link-prediction over ogbn-arxiv) — not the synthetic A(t) transforms of ADR-200.

Outcome ADR: ADR-202 — WIN. Gate was pre-registered and frozen before any contender run (docs/plans/bet1-productionize/PRE-REGISTRATION.md), per the prove-not-hype protocol.

Result (WIN at n=20k and n=50k)

policy (n=20k)	recall@10	cumulative rebuild cost	evals/q
B always-rebuild	98.7%	246s (30 builds)	982
A reuse-only	94.0%	0s	1034
P k=4 (shippable)	98.7%	59s (24%)	983

• Reuse transfers in-regime: pure ReweightOnly holds within 2% recall@10 of full rebuild up to a 40% top-10 churn ceiling — identical at n=20k and n=50k, at/beyond ADR-200's synthetic ~36%. At low churn reuse is occasionally above rebuild.
• Periodic recovers the high-churn tail completely: Periodic{k:4} gap −0.01% (20k) / above rebuild (50k) at 20–24% of rebuild cost, equal per-query work. ADR-200's hybrid finding reproduced on real drift.
• Teeth: the stale control collapses (92%→33% at 20k), proving the benchmark is genuinely drift-sensitive.
• Honest caveat (reported, not hidden): pure reuse run past its ceiling on a deliberately overdriven trajectory decays (−4.73% to 67% churn, 1.05× evals) — which is exactly what the periodic policy is for. The shippable periodic policy carries neither penalty.

What's in the PR

• crates/ruvector-diskann/src/reuse.rs (feature reuse-under-drift, default off → shipping build byte-identical): RebuildPolicy{AlwaysRebuild, ReweightOnly, Periodic{k}} + DriftingIndex. The index owns only the rebuild decision; the consumer (GNN) owns the drifting embeddings and passes snapshots to on_metric_update/search. Native reuse hook — greedy_search already takes vectors externally. 5 unit tests.
• crates/ruvector-gnn/examples/diskann_real_trajectory.rs: the validation harness (trajectory generator + 4 contenders + gate eval). Embeddings live on the unit sphere so the contrastive metric and the diskann L2 ranking agree.
• docs/adr/ADR-202-*.md + docs/plans/bet1-productionize/PRE-REGISTRATION.md.

Honesty notes

• GraphMAE::train_step takes &self and does not update weights — it cannot produce drift. The trajectory is built from the repo's genuine learnable primitives (Optimizer + info_nce_loss + SGD on embeddings), documented up front.
• The frozen gate's "within 2% over the early trajectory" clause is operationalized as the holding ceiling (max contiguous churn where reuse stays within 2%) — the regime-resolved statistic the gate named, not the trajectory-wide mean (which deliberately overdrives past the ceiling). Both statistics are reported.

Self-contained off main; depends only on ruvector-diskann + ruvector-gnn. Independent of #535.

Refs #534