perf: qualify range and index roadmap experiments

[josephjohncox]

Summary

Runs the six-area performance roadmap as bounded, evidence-gated experiments and lands only the changes that passed their fixed, pre-registered gates. Failed runtime candidates are deleted rather than shipped behind loosened thresholds. Adds concrete ExactBatchRangeSet and BoxIndex examples, and hardens every experiment verifier against forged evidence.

The canonical stable API (RangeSet, create_range_set, BackendRegistry), exact-batch geometry, payload algebra, atomicity, and all existing gates are unchanged.

Accepted into the runtime (2)

• Geometry-only RangeSnapshot cache (treemendous/rangeset.py): reuses one immutable snapshot while geometry is unchanged and no payload policy is set; payload-bearing snapshots still clone/detach exactly as before. 40-block confirmation: unchanged N=10 000 upper-95 0.00057 (≤0.25), cached N=10000/N=1000 median 0.958 (≤1.50), worst write-then-observe upper-95 1.035 (≤1.10). Scalar-mutation regression vs. immutable baseline: median 0.98 (100 samples).
• Lease projection caches + O(1) fence lookup (treemendous/applications/_shared/leasing.py, leasing/_common.py, Candidate B): direct token lookup with zero public snapshots and zero linear scans; lazy immutable projections invalidated on publication. 30-block: fence upper-95 0.177, repeated-snapshot N=8192 upper-95 0.266 (both ≤0.50).

Rejected and removed (deletion, not gate-weakening)

Area	Result	Reason
RangeSet.mutate_many transaction (E2)	REJECTED	B16/N1000 upper-95 >1.07 per payload, 23× with no payload; BackendAdapter has no atomic state-replacement primitive. No method/protocol/export added.
Segmented COW native storage (E3)	REJECTED	Wins large-N (N=100k/B16 local <0.60×) but tuned one-block fast path still gives N64/B16 upper-95 1.116 > 1.10. Runtime fully reverted to vector storage.
Concrete-engine backend injection (E5)	REJECTED	cpp_boundary confirmed faster, but current-default py_boundary upper-95 hit 1.125 on pool_group (≤1.10 default-stability gate). No injection retained.
Lease Candidate D (authoritative free-delta)	REJECTED	N8192/N2048 normalized time 3.7× (≤2×). Reverted.
Adaptive BoxIndex for RadioSpectrum (E6)	REJECTED	Projection wins narrow queries (0.80–0.91×) but retained memory 1.30–1.37× (≤1.25×) and broad queries regress to 1.24–1.38×. Runtime stays linear BoxIndex(2).

New executable examples

• ExactBatch: atomic_memory_map_updates, atomic_partition_availability_updates, genomic_mask_batch_updates
• BoxIndex: warehouse_space_time_reservations (3D), video_region_timeline_overlap (3D), robot_volume_time_conflicts (4D)

All are registered in examples/README.md and executed from an unrelated CWD by tests/docs/test_readme.py.

Diagnostic evidence (informative only)

• Exact-batch application matrix (180 cells): ExactBatchRangeSet beats scalar cpp_boundary replay from B=1 at N≤1000 across all shapes, but loses at N=10 000 (1.1–3.6×) — the O(N) full-state copy E3 tried and failed to remove.
• Segmented-storage rejection is durably reproducible from a tracked patch (tests/performance/experiments/fixtures/exact_batch_segmented_tuned.patch) plus an archived triplet under docs/evidence/experiments/, bound by source/binary/patch SHA-256 and verifiable offline (just verify-exact-batch-storage-archive).

Verifier hardening

Every experiment verifier rejects fabricated runtime/build/backend provenance, bool/int and int/float coercion, duplicate keys, non-canonical JSON, and relabeled/duplicated matrix rows. The snapshot benchmark's order-confounded pooled analysis was replaced with pre-registered balanced AB/BA blocks and whole-block bootstrap.

Boundaries preserved

• No compatibility façade, deprecated alias, silent fallback, or weakened exact result.
• Experiment code lives only under tests/performance/experiments/; no stable package API, ABI, or gate changed.
• Explicit backend/index/guard qualification errors are never caught to select a fallback.

Validation

• just check (layout, scenarios, Ruff, format, mypy, full pytest + coverage, packaging/docs, bytecode) passed locally.
• Focused experiment suites: 166 passed; snapshot-cache + lease-publication + exact-batch suites green.
• All 12 tracked examples and both README quickstarts execute from an unrelated CWD.
• Durable archive verifier, snapshot/lease/application-backend generators, and their strict verifiers round-trip.

Experiment matrices are machine-specific diagnostics; accepted-runtime decisions rest on the fixed gates above, not on any single headline number.