feat(proximity): add native TurboQuant accelerator

[forhappy]

Summary

• Add an independently implemented native TurboQuant-MSE proximity accelerator with deterministic structured rotation, canonical 2/3/4-bit encoding, exact authoritative reranking, sync/async execution, proofs, content walking/copy/GC, bounded caches, and all maintained language bindings.
• Keep Auto disabled and expose TurboQuant only through explicit backend selection while mandatory release gates remain open.
• Correct L2, cosine, and inner-product reconstruction-norm scoring and fuse dot/norm accumulation into one packed-code traversal.
• Correct the qualification harness to use an authoritative exact ProximityMap search over persisted canonical vectors. Benchmark schema 4 rejects all older recall evidence.

Correctness and verification

Passed locally:

• cargo fmt --all -- --check
• git diff --check
• cargo clippy --all-targets --all-features -- -D warnings
• complete all-feature Rust suite and 74 doctests
• 23 TurboQuant integration tests and 24 focused algorithm tests
• 17 qualification runner/summarizer tests
• bounded decode/lifecycle fuzz smoke, WASM/browser-WASM, Miri, and maintained binding suites
• schema-4 exact-resume and digest validation for all eleven retained cells

Schema-4 evidence

Eleven focused cells at implementation revision 729c2ffa cover 10K/100K × 768, all metrics at four bits and 8× reranking, matching 16× cosine/product diagnostics, and exhaustive 10K cosine.

Records	Metric	TQ recall 8×	PQ recall 8×	TQ/PQ p95
10K	L2 squared	1.00	1.00	6.23×
10K	Cosine	0.00	1.00	2.90×
10K	Inner product	1.00	0.00	8.01×
100K	L2 squared	1.00	0.00	9.54×
100K	Cosine	0.00	0.00	7.36×
100K	Inner product	0.00	0.00	9.48×

Exhaustive 10K cosine returns 1.00 recall for both implementations. A qualification-scale Rust regression proves the default failure is unavoidable under the frozen contract: the authoritative exact top 10 all share one packed code with 222 lower-key records, while the shortlist cap is 80 and ties order by key. The approved 16× window still returns 0.00 at 10K; 100K cosine and inner product reach 1.00 at 16×.

TurboQuant builds faster than PQ in every sampled row, but search and sidecar-size gates fail. Full details and raw artifacts are in the schema-4 report.

Release status

This PR remains a draft. Open gates are:

• approved resolution for the incompatible 10K cosine packed-code/shortlist contract
• complete 45,360-cell single-revision matrix
• pinned-host qualification
• repository-wide binding inventory with 2,998 pre-existing incomplete mappings
• legal/patent release disposition

The feature-specific proximity and fuzz checks pass. Current unrelated DynamoDB/S3 workflow failures are inherited repository issues and are not hidden by this PR.

[forhappy]

Qualification update (0d1b3a9): fixed the cross-target conformance failure by replacing host-libm-generated fixture inputs with frozen IEEE-754 bit inputs; added an independent dense Gaussian-QR development oracle with retained 128/200/768/1536/3072 × 2/3/4-bit distortion and recall evidence; and added ordinal fault injection over every discovered TurboQuant publication, cold verification, forced search/rerank, and proof/closure store boundary. Focused tests and warnings-denied Clippy pass locally. Auto remains intentionally disabled and production-scale qualification/legal gates remain open.

[forhappy]

Retained qualification update (379463d): corrected accelerator sidecar accounting to exclude referenced authoritative source objects; regenerated the 1K × five-dimension × three-metric default-path runs from committed harness revision d734772; and checked in the raw CSVs plus a bounded-tier report. All 15 TurboQuant recall@10 rows are 1.0 and match PQ, but TurboQuant search and derived size are worse on this fixture, so Auto remains disabled. Production-scale, cross-host/release-command, and legal gates remain open.

[forhappy]

Performance and conformance update: 60dfaa7 replaces per-coordinate generic unpacking with canonical width-specialized 2/3/4-bit decoding. On the retained 1K L2 arm64 tier, scalar scan improved about 33–44% across dimensions (36–37% at 768/1536) with unchanged 1.0 recall. c701192 refreshes the raw evidence, and e3cd7df adds a browser-WASM test that reproduces the checked-in Rust descriptor and TurboQuant manifest CIDs. Auto remains disabled because current scan/size results still do not meet its comparative gates.