v1.4.1 — docs, reproducibility, positioning + estimate_storage fix

Docs / reproducibility / positioning release, plus one real bug fix. No breaking library API changes. Published to PyPI 1.4.1.

Fixed

• PCAMatryoshkaPipeline.estimate_storage() no longer reports hard-coded 1024→384 @ 3-bit regardless of configuration — now an instance method defaulting to the pipeline's real input_dim/output_dim/bits (static estimate_storage_for(...) keeps the dimension-agnostic form). Regression-tested. (External review #3.)

Added

• benchmarks/canonical_embedding.py — one tested harness running every embedding method (flat / PQ / OPQ / IVFPQ / RaBitQ / PCA-only / TQ-only / PCA+TQ / ADCIndex) at an identical rerank protocol on public ann-benchmarks data.
• notebooks/claims/ — one Colab notebook per evidence-ladder claim (00 canonical SOTA table; 01–04 embedding/CPU incl. an OOD anisotropic/heavy-tailed stress test; 10–12 KV-cache/GPU).
• CLAIMS.md — claim → reproduction table; docs/claims.md evidence ladder; docs/api-stability.md stability tiers; benchmarks/RESULTS_canonical.md protocol.

Changed

• README: Version block, slimmed headline (claims → CLAIMS.md), inline API-stability table, "Not to be confused with" (incl. vLLM's TurboQuant integration), one-Run all SOTA pointer. Static test count → CI badge.

Addresses three rounds of external review (evidence organization, positioning/version consistency, and the technical estimate_storage fix + OOD robustness).