warm-python 0.6.4__tar.gz

warm_python-0.6.4/CHANGELOG.md

@@ -0,0 +1,562 @@
+# Changelog
+
+All notable changes to this project are generated from the commit history.
+Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/) +
+[Conventional Commits](https://www.conventionalcommits.org/).
+
+[v0.6.3 migration and integrity notes](https://writerslogic.com/warm/docs/latest/releasing/v0.6.3/).
+## [0.6.4] - 2026-10-28
+
+### Added
+- Add optional AML provenance and evidence research tools
+
+### Changed
+- Organize bindings and research source modules
+
+### Documentation
+- Align landing page with retrieval architecture
+- Add Glama score badge (#316)
+- Explain crate recovery environment policy (#322)
+- Explain npm registry propagation gate
+
+### Fixed
+- Bootstrap Python docs from the current source layout
+- Synchronize release and OpenAPI versions
+- Preserve guarded final benchmark allocations
+- Enforce provenance semantics and cached status expiry
+- Preserve Mem0 recovery and completion integrity
+- Clarify MCP tool behavior for Glama (#317)
+- Reduce crate archive for crates.io (#321)
+- Verify every npm platform before install (#320)
+- Verify every npm platform before install
+## [0.6.3] - 2026-10-26
+
+### Added
+- Add reproducible AML GCE installer
+- Add persistent AML host launcher
+- Add AML Cycle 2 adapter and query repair
+- Make MCP server scorable by Glama (#313)
+- Add a blocked learned-sparse inner-product index (big-ann sparse track) (#294)
+- Scale query-aware projection to millions of learn queries (#295)
+- Add label-filtered exact uint8 search with per-label inverted files (#291)
+- Add direct IP serving and independent candidate retention (#207)
+- Add compact adjacency and publish ANN cost results (#206)
+- Assemble bounded conversation context with exact source provenance (#203)
+- Add optional query-aware graph routing (#204)
+- Improve memory recall within a small round budget (#202)
+- Diagnose streaming recall and verify crash boundaries (#201)
+- Add exact Boolean lexical search and blocked postings (#200)
+- Batch inserts and a Python UpdatableIndex for the updatable vector graph (#190)
+- Count matching documents and ingest unstemmed terms (#193)
+- Persist and update mapped vector graphs (#181)
+- Preserve original conversation evidence and audit memory quality (#184)
+- Add VIBE diagnostics and rotation controls (#183)
+- Euclidean and inner-product metrics in the Python VectorIndex (#186)
+- Clustered private ranking of dense codes at n = 1280 (#177)
+- Server-mode comparators and an warm-server row in the harness (#157)
+- Raise WarmError with the server's error code and message (#145)
+- Add NGT, ScaNN, PyNNDescent, Voyager and DiskANN to the ANN harness (#156)
+- Euclidean metric for the vertex index and SIFT in the harness (#150)
+- Expose the vector index to Python and add ann-benchmarks and VIBE wrappers (#149)
+- Framework adapters for agent and RAG stacks (Python and JS) (#106)
+- Stable C ABI with Go, Java and .NET bindings (#109)
+- Index-routed server vector queries, deferred ingest modes, memoized encoder, and round-2 research kill tests (#139)
+- Local embeddings (pure-Rust engine, pinned model packs, LocalEmbedV1) (#132)
+- Evidence manifest, fetch script and worktree audit record (#117)
+- Format verifier, golden stores for every format, drift gate, crash-window fixes (#126)
+- WebAssembly build of the document store (@writerslogic/warm-wasm) (#129)
+- Production warm-server (scoped auth, observability, TLS, snapshots, deploy artifacts) (#127)
+- UniFFI mobile bindings (Swift, Kotlin/Android) (#121)
+- Warm-mcp MCP server, launchers, memory-tool backends and editor integrations (#110)
+- Memory API with temporal validity, supersession, explain and opt-in policies (#114)
+- Warm-import, one-command migration into WARM (#108)
+- Platform kernels (AVX512-VNNI, NEON fallback) with kill-test results (#107)
+- Integrate the parallel sessions' document store, compaction and server work (#81)
+- Document store with compact postings and file-backed text, staged compaction, multi-tenant server and the evidence client
+- Learned-sparse impact index served by WARM, measured against Anserini
+- One series per patience value in the ANN comparison tables
+- Stemmed BM25 and min-max score blend in the document API
+- Patience stop rule, opt-in 1-bit screen and error-returning builds for the vertex graph index
+- Competitor ANN harness for SymphonyQG, RaBitQ, NGT and Lucene HNSW
+- Check the graph cache header and verify the merged qgraph defaults
+- Add integer vector stores and per-vertex codes to qgraph
+- Batch Qwen3 decoding and padded encoder batches on Metal and CUDA
+- Add per-vertex 8-bit code graph index with residual re-rank
+- Add held-out tuning mode to public-bench ann-qgraph
+- Add on-device Qwen3 model stages behind local-models
+- Add quantized graph ANN index and its public benchmark mode
+- Add warm-server HTTP service for the Python SDK
+- Subset-union bundle experiment and capacity sweep
+- Add experimental query-private similarity scoring behind private-search
+
+### Changed
+- Clean up memory API, provenance, security and core modules (#214)
+- Clean up the full-text storage engine (#213)
+- Clean up bindings, server and MCP (#212)
+- Clean up qgraph kernels, projection and update format (#211)
+- Prune uncited experimental research code (#210)
+- Consolidate audit entry serialization (#185)
+- Drop rejected intersection kernel candidates
+
+### Documentation
+- Prepare npm listing and runtime release (#318)
+- Surface holographic operations in search
+- Note npm and runtime release
+- Prepare npm listing and runtime release
+- Document Glama release and score verification
+- Polish MCP package pages and link website (#311)
+- Make source quickstart accurate
+- Refresh branding and improve integration and search guides (#308)
+- Drop host names and local install notes, fix the open-test citation (#221)
+- Replace private paths and hosts, route large result JSON to LFS (#217)
+- Rewrite README with sourced benchmark badges and current architecture (#215)
+- Serve the documentation at docs.writerslogic-warm.com (#189)
+- List every Cargo feature and correct the server and MCP rows (#173)
+- Build the rustdoc with warnings denied in CI (#148)
+- Remove internal scoreboards, session records and process notes (#140)
+- Remove session logs, handoffs and planning files (#137)
+- Embeddings default-build check and handoff update (#135)
+- Render OpenAPI boolean property schemas (#131)
+- Passage text is encrypted at rest; document the text segment layout (#115)
+- Redesigned site with landing page, generated API reference, search and browser checks (#125)
+- Persisted file format specification and stability policy (#119)
+- Versioned site, quickstarts, offline agent demo, README, contributor guides (#105)
+- Ideas tree for October 2026 with its 10-minute proxies
+- Preliminary nytimes VGraph comparison on the M4 and the next-session specs
+- Record the green CI and the clippy toolchain for the gate
+- Record that main is pushed and in sync
+- Record that all work is on main and what was superseded
+- Record the merged qgraph state and what is unverified
+- Record the integration state
+- Standardize README header (#71)
+- Update continuation state
+- Record the index and model-stage outcomes
+- Record the holographic prototype outcome
+- Record continuation state for the next session
+- Plan the holographic fact memory and its tests
+- Plan local model stages inside WARM
+- Define the single-run protocol for LongMemEval_M
+- Set the speed and accuracy targets and record the compute limits
+- Plan the LongMemEval_M run with larger models
+- Record the encoder test and the agent-memory benchmark plan
+- Attribute semantic search to the document API
+- Update changelog [skip ci]
+- Record the outcomes of the algorithm work
+- Give the SDK README the project header, badges and a quick start
+- Use absolute logo URLs so the README renders on package registries
+- Update changelog [skip ci]
+
+### Fixed
+- Keep Python source archive within PyPI limit (#314)
+- Wait for MCP package indexes before registry publish (#312)
+- Rename Python packages and polish PyPI pages (#310)
+- Persist ANN benchmark indexes before serving
+- Update benchmark harnesses for current Clippy
+- Complete graph search results after component exhaustion (#301)
+- Refuse web-page and rebound-host requests on an unauthenticated server (#285)
+- Check checkpoint table sizes, stage model packs and semantic indexes privately, keep oversized fact turns out of range (#281)
+- Reject damaged delta offsets, save vertex indexes under unique temp names and assert re-rank row lengths (#280)
+- Domain-separate Sigstore signing and close KERI, CAWG, revocation and DID verification gaps (#279)
+- Scope tenant metrics to the key, clean up abandoned restore uploads and stop insight growth (#278)
+- Sync migrated and snapshot copies and the API-key directory before reporting success (#276)
+- Normalize extreme angular rows, bound graph-cache reads and stage Python deletes all or none (#275)
+- Persist revocations before applying them and add a strict status-list mode (#273)
+- Keep documents and format scans consistent after failed rebuilds and malformed frames (#272)
+- Keep tenant generations unique across reopen and harden tenant lifecycle (#265)
+- Leave the journal unchanged when a commit sync or segment rotation fails (#263)
+- Apply a journaled batch in full when one mutation fails (#264)
+- Domain-separate batch provenance Merkle trees and version the signed root (#243)
+- Decode sigstore base64 strictly and write the key event log atomically (#257)
+- Bind provenance signatures to the record they verify (#242)
+- Track VIBE job CPU per process and kill orphaned config processes in the Modal runner (#232)
+- Update ctutils past RUSTSEC-2026-0385 (#234)
+- Build the updatable index with alpha 1.0 by default (#229)
+- Read vector result text under the store lock (#227)
+- Mark the gzip LFS exception as binary (#225)
+- Read search result text under the mutation gate (#223)
+- Core review fixes for provenance, memory API and server (#219)
+- Measure one durable ANN mutation per operation (#180)
+- Make a purged store match one that never held the document (#179)
+- Durability-fair server verdicts and harness adapter checks (#175)
+- Match server durability on macOS and build ScaNN with the dataset's metric (#171)
+- Score docs search on the answer to the full query (#170)
+- Judge recall status from the index of the revision searched (#168)
+- Modal shard runner, text harness defects and the x86 text results (#152)
+- Make each journal commit durable on macOS with F_FULLFSYNC (#151)
+- Name the passphrase when an encrypted store cannot be decrypted (#144)
+- Strip debuginfo only so macOS 27 can load the native addon (#136)
+- Clear code-scanning alerts and restore the server_requests fuzz target (#134)
+- Zero replaced document text in the text segment (#128)
+- Purge physically erases deleted text (#124)
+- Encrypt the document text segment at rest (#111)
+- Windows compaction rename, changelog drift, dependabot npm 404 (#112)
+- Check tenant directories against the canonical tenants root
+- Fail open on arena corruption, checksum the manifests, zeroize the AES key schedule
+- Scrub 256 KiB of stack below the key derivation and locate surviving key copies in the scan test
+- Gate scrub_stack and the AES cipher impl behind the security feature
+- Satisfy clippy 1.99 lints in the qgraph kernels and graph cache
+- Keep greedy decoding at batch 1 unless a batch is set
+- Reserve Modal wave cost before launch so preemption cannot bypass the cap
+- Run WARM model stages on L4 on Modal
+- Decode benchmark data with as_chunks for the current clippy
+- Serve SDK queries by exact cosine through the document API
+- Resolve leftover merge markers in .bestpractices.json
+- Return query results best-first on exact-scan, multi-shard and federated paths
+
+### Performance
+- Accelerate lexical query execution and decoding
+- Improve phase memory and recover conversation premises
+- Add opt-in vector search improvements and memory controls
+- Reduce filtered-search allocation and scoring overhead
+- Share immutable position sources by document slot (#305)
+- Add an opt-in bound-gated exact re-rank to the vertex graph search (#296)
+- Resolve phrase positions through posting ordinals (#290)
+- Decode packed posting blocks with AVX2 (#293)
+- Skip frontier searches for noncompetitive delta neighbours (#289)
+- Spill delta rows during vector consolidation (#286)
+- Release old graph pages before delta insertion during consolidation (#284)
+- Score construction candidates four at a time and with VNNI on AVX-VNNI cores (#283)
+- Avoid copying surviving vectors to replace a deleted graph entry (#282)
+- Reduce maintenance allocations and fix retrieval regressions (#274)
+- Skip inverted-index hits no better than the worst kept one (#277)
+- Prune lexical top-k with block-max MaxScore and best-blocks-first term scans (#237)
+- Merge the delta into the streaming base instead of rebuilding it (#271)
+- Reuse warm ingest interners across rayon splits and batches (#268)
+- Return inverted-index hits as slot numbers instead of decimal strings (#267)
+- Score Hopfield queries against the stored patterns in place (#266)
+- Cut Python SDK per-call overhead (#269)
+- Fuse the first three Walsh-Hadamard stages per block of eight (#260)
+- Sign batch provenance credentials in parallel (#258)
+- Keep the pool's expanded flags in a bitmask (#252)
+- Clear the streaming delete-runbook baseline and drop the insert row copy (#238)
+- Count majority bundles in a per-thread histogram (#259)
+- Count majority-vote encoder levels at the touched indices (#256)
+- AVX-VNNI traversal kernels with the query sum computed once per query (#251)
+- Huge-page advice on by default, with counters and a refusal fallback test (#241)
+- Track visited vertices in a bitset cleared by touched word (#240)
+- Reuse construction scratch buffers across insertions, merges and fills (#248)
+- Prepare single document writes with the store's kept interner (#254)
+- Build chunk ids at their exact size and check them without allocating (#253)
+- Reuse one output buffer when recovery tests a damaged journal tail (#250)
+- Rebuild a checkpoint without the postings of its dead documents (#249)
+- Single-thread vertex search (prefetch terms, SIMD re-rank, 8-bit codes in the harness) (#226)
+- Open text stores from a durable document checkpoint (#235)
+- Persist the warm-server vector graph and open it at start (#228)
+- Qgraph review fixes (#220)
+- Full-text engine and storage review fixes (#218)
+- Eliminate redundant graph construction and query work (#209)
+- Reduce posting block allocation overhead (#208)
+- Bound streaming consolidation memory with mapped rows (#205)
+- Rank unfiltered lexical queries in the slot table at any store size (#197)
+- Store construction adjacency in one fixed-slot array (#182)
+- Faster single-threaded text ingest (#178)
+- Lean BM25 search endpoint and in-place fact replacement in warm-server (#176)
+- Build the vertex graph faster (#174)
+- Serve the vertex index from a memory-mapped file (#172)
+- Binary journal frames, in-place posting deletes and an ingest word interner (#155)
+- Rank unfiltered lexical queries in a per-thread slot table (#154)
+- Round-3 ingest, memory and ANN work with research results (#143)
+- Quantized construction search for the graph build; cache header records it
+- Keep the default neighbour loop free of the delta decoder and the cap counter
+- Quantized construction search and owned-input build for the proximity graph
+- Neighbour cap for the per-row search, lint fixes, x86 paired rows
+- Byte visited marks for the unfused block path; close the neighbour-block mechanism
+- Neighbour-block fast-scan codes with evidence fusion and a certified exit for the vertex graph
+- Unbiasing scale for block codes, sorted candidate pool, delta-coded adjacency
+- NEON per-row 1-bit screen kernel with its microbenchmark kill test (screen stays off)
+- Screen exact distances in vertex graph construction with 8-bit bounds
+- Add 4-bit traversal codes, aligned rows, BFS reorder and 3-byte ids to the vertex index
+- Measure Metal batching and set per-stage defaults
+- Prefetch rows ahead in qgraph fill, prune inputs and edge encoding
+- Use a bitset visited set and prefetch adjacency in qgraph construction
+- Scan both planes of 2-bit qgraph edge codes in one NEON pass
+- Speed up qgraph construction without changing the built index
+- Add FastScan edge kernel, batched pool merge and 2-bit qgraph edge codes
+- Gallop before AVX2 on skewed sizes; add intersection kernel bench
+- Speed up from_dense with branchless term compaction and add encoder-eval
+- Route sparse queries to the exact inverted index
+
+### Bench
+- Sync the VIBE fork and current engine pins (#307)
+- Record ordered payload rank prototype evidence (#306)
+- Record frozen bounded-consolidation tradeoffs (#303)
+- Record rejected consolidation scratch reuse (#302)
+- Verify PGO gains with an uncontended paired repeat (#300)
+- Record the rejected low-rank OOD experiment (#299)
+- Add Gorder window locality diagnostics (#298)
+- Record quantized delta heldout recall rejection (#297)
+- Record the neighbour-major block kill test against the per-row path (#292)
+- Build the VIBE image for x86-64-v3 and pre-register the OOD grid (#287)
+- Attribute the streaming runs to the measured commit (#231)
+- Add an WARM backend for Mem0's memory-benchmarks harness (#262)
+- Measure memory API write costs and record host-layer findings (#245)
+- VIBE wrapper with 8-bit codes, query-aware OOD builds and huge pages (#255)
+- Add a LoCoMo captions variant measured against the matched protocol (#261)
+- Profile vector-index build allocations and vertex-order locality (#247)
+- Count heap allocations per write-side phase of the document store (#246)
+- Add LoCoMo retrieval round 2 diagnosis and gated packing levers (#239)
+- Add LoCoMo results (#236)
+- Add a batch mode for the LongMemEval reader and judge (#233)
+- Run llama-128-ip and yi-128-ip on Modal with mlann-rf and roargraph arms (#230)
+- Add a crash-safe Modal runner for the VIBE float datasets (#224)
+- VIBE adapter for euclidean and ip datasets (#194)
+- Add single-core ann-benchmarks harness results on the Core Ultra 9 (#198)
+- Run the Bemula forgetting benchmark against the memory API (#199)
+- LoCoMo adapter and partial results (#196)
+- Score LongMemEval dev under the official reader and judge protocol (#195)
+- Pin the ann-benchmarks image and run euclidean datasets (#187)
+- Add Oct 21 2026 Core Ultra 9 competitive results (SIFT-1M, BEIR text) (#158)
+- Reproducible comparison harness and self-updating table (#122)
+- Per-session quota on the tuned LongMemEval turn ranking fails on cached dev scores
+- Add grid-matched vertex search pairs and correct the alignment claim
+- Record paired held-out timing of the vertex search encodings
+- Narrow the FAISS comparison and vendor the clean edge baselines
+- Record paired held-out timing of the per-vertex code index
+- Add paired-timing sync to ann-qgraph
+- Record held-out recall and memory of the per-vertex code index
+- Record timed held-out glove run of i8 vertex-code qgraph
+- Correct reference eval counts in memory-first results
+- Record held-out recall screening of qgraph memory layouts
+- Record the s-dev-warm cost overrun and stop
+- Record held-out qgraph tuning on nytimes
+- Measure WARM fact extraction against the Python stage on S dev
+- Add seeds and an equal-bytes index control to the holographic prototype
+- Record quantized graph results against HNSW
+- Prototype the holographic fact memory and measure it on LongMemEval_S dev
+- Make the Modal cost cap cumulative across relaunches
+- Key LongMemEval fact extraction by content and re-tune on dev
+- Record the LongMemEval dev results and the Modal dry run
+- Tune the LongMemEval pipeline on dev and record the ablation
+- Add the LongMemEval multi-key pipeline, Modal job and dev sweep
+- Freeze the LongMemEval 100 dev / 400 held-out split
+- Add LongMemEval_S retrieval baseline (dense, lexical, hybrid)
+- Add public BEIR and ann-benchmarks comparisons
+- Report encoder candidate recall after exact rerank
+
+### Deps
+- Rand 0.10, rand_chacha 0.10, getrandom 0.4 for bindings/wasm (#120)
+
+### Fuzz
+- Targets for every untrusted-input parser (#116)
+
+### Merge
+- Include AML GCE installer on main
+- Include persistent AML host launcher on main
+- Include AML Cycle 2 adapter on main
+- Integrate vector search improvements
+
+### Research
+- Harnesses, result files and notes from the parallel sessions
+- X86 confirmation of the kernel fix and the build lever on Modal (#80)
+- X86 confirmation of the kernel fix and the build lever on Modal
+- Superiority session, October 16 to 17 (#76)
+- Record the V37 compact-postings serving row and the session handoff
+- Accept V40 serving row, record the V38 transport and 2C relevance measurements
+- Security and privacy matrix harness and first matrix
+- Durability matrix harness and first matrix against SQLite and Lucene
+- Checkpoint native harness validation draft
+- Record native benchmark gate checkpoint
+- Establish checked LongMemEval dev answer results
+- Checkpoint final local reader batch and faster validation
+- Preserve dev reader progress and speed pilot audits
+- Checkpoint local reader evidence and judge progress
+- Checkpoint judge instability and local evidence
+- Checkpoint dev answer scores and repeat study
+- Freeze dev reader protocol and checkpoint quota blocker
+
+### Style
+- Address code quality findings in the integrated harness scripts
+- Use as_chunks for the fixed four-byte float and id decoding in the ANN bins
+- Read cpuinfo once and drop an unused import
+- Address code quality findings in the benchmark scripts
+- Close files with pathlib and unpack the tamper case explicitly
+- Rustfmt
+## [0.6.1] - 2026-10-05
+
+### Added
+- Introduce IdentityRegistry enforcing W3C DIDs, Verifiable Credentials, and POSME Receipts for queries
+- Add AsyncWarmClient, batch uploading, and metadata filtering
+- Add Pydantic AI and SmolAgents tool wrappers
+- Add Phidata integration and fix broken Rust examples
+- Add Semantic Kernel and DSPy integrations to Holographic SDK
+- Add LlamaIndex SDK adapter and CI pipeline
+- Pivot to Zero-Trust Edge Architecture (FHE-Lite, AGPL-3.0, SIMD)
+- Python bindings + PyPI packaging for holographic-vsa (maturin/pyo3)
+- PhaseResonator reusable-index API + clean re-export + doctest (§20 capability)
+- PhaseHVec core type + real Frady/Kent resonator on quantized-phase substrate
+- Counting membership store with Poisson z-score readout (4x Bloom capacity)
+- Real SHA-256 commitment for verifiable phasor mutation (lane 3)
+- Multi-hop reasoning over the phasor memory (retrieve_path)
+- PhaseGraph -- phasor relational memory with relation algebra + retrieval
+- Presence-gate the observer effect (no confabulation-on-query)
+- Persist ConnectionGraph via its event log; survives restart
+- ConnectionGraph -- plastic event-sourced relation store, wired into engine
+- Trust-anchored provenance verification and sharded ANN persistence
+- Runnable cross-verification of the cogmem C2PA sample
+- Add provenance system with COSE, VCs, C2PA, JUMBF, Sigstore, KERI, CAWG
+- Multi-scale encoding, morphological decomposer, fault-tolerant federation
+- Add research-grade benchmarks, scaling analysis, and visualizations
+- Add sparse Clifford algebra multivector type (CliffordVec)
+- Add Hopfield-Fenchel-Young energy-based associative retrieval
+- Define HolographicAlgebra trait boundary for future geometric algebra
+- Improve SOTA for analogy, planner, and concept synthesis
+- Add cost/total_cost fields to planner N-API types
+- Evidence-weighted multi-hop confidence, ranked results
+- Expose cleanup_vector N-API method for standalone Hopfield denoising
+- Add agency layer (GoalStore, Planner, QuestionGenerator, SelfModifier)
+- Add cognition layer, distributional refiner, CI coverage
+- N-API bindings, persistence, load_from_log, compact for meaning memory (G6)
+- Integrate meaning memory into WarmCore, wire write/delete paths (G5)
+- Fuzzy_structural_query + multi_hop_query pipelines (G4)
+- AtomMemory, CompositeMemory, TripleStore, RuleStore, Decomposer (G3)
+- Shared utilities + IndexedMemory with Hopfield attractor (G1-G2)
+- Rename to writerslogic-warm, fix npm logo, update all URLs
+- Graph engine with multi-hop traversal, inference, temporal, federated queries
+- Complete security integration, DP wiring, docs, LF readiness
+- Add security features behind feature flag for LF Decentralized Trust readiness
+- Migrate to writerslogic org, add npm publish pipeline
+- Eliminate remaining gaps, expose train_nsg/train_ivf via N-API
+- Expose diffusion config through WarmConfig and N-API
+- Configurable thresholds, N-API config constructor, status getters
+- Add query_sequence, elevate test coverage for weak features
+- Add multi-shard support with auto-sharding
+- Initial release of WritersLogic Adaptive Retention & Memory (WARM)
+
+### Changed
+- Centralize SDK HTTP client and add CrewAI + Embedchain platforms
+- Consolidate LangChain, LlamaIndex, and Haystack adapters into unified warm-sdk
+- Extract shared wire codec (write/read_lp_str, write/read_deltas)
+- Remove dead self-inverse resonator; move ResonatorConfig to phase_resonator
+- Extract meaning_ctx() helper shared by structural_query and multi_hop
+- Gate experimental VSA modules behind experimental feature
+- Migrate WARM agent identity from did.cose to CAWG ICA
+- Deduplicate RNG pruning, unify patterns, add doc comments
+- Eliminate dead code, add delete/compact with persistence
+- Remove redb dependency, use in-memory FxHashMap with arena log persistence
+
+### Documentation
+- Update changelog [skip ci]
+- State the AGPL-3.0-or-later license in the security policy
+- Record algorithm findings and private-search threat models
+- Update changelog [skip ci]
+- Rewrite README to highlight Zero-Trust Enterprise Security and Python SDK
+- State the testing policy explicitly in CONTRIBUTING.md (#66)
+- Sync changelog (#54)
+- Align guidance and track deferred production work (#52)
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Standardize badges (add OpenSSF Scorecard; consistent order/format)
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Overnight closeout — queue exhausted, 337 tests green, flag pre-existing §21 bin lint
+- Retract 'info-theoretically impossible' framing; 0.1D Plate is a readout artifact, closing gap to ceiling is open (§22)
+- Block codes also floor (§24); capacity gap is retrieval-vs-information, open not impossible
+- DETERMINISTIC-RESONATOR.md — §20 validation write-up from in-loop 24-seed run
+- Reproducibility header for resonator-factorize (fixed seeds, correct repro command)
+- Correct resonator-factorize repro command — bin is std-only, no experimental feature
+- Wire deterministic-resonator capability into RESEARCH.md (discoverability + correct citation)
+- §29 confirmed on unseen seeds — 4-bit matches float, 8-16x footprint win; log floor-exploration findings
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Record why presence gate uses an absolute margin, not SNR
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Pre-register non-self-inverse binding + nonlinear readout experiment
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Replace logo with continuous-rotation animation
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Restructure README with collapsible sections
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Rewrite README — fix logo for npmjs, restructure with quick start first, improve clarity
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Update changelog [skip ci]
+- Add agent-provenance stack cross-reference to README
+- Update changelog [skip ci]
+- Update all documentation for v0.5.0 accuracy
+- Add meaning memory architecture, structural queries, attractor cleanup
+- Add meaning memory architecture, structural queries, attractor cleanup
+
+### Fixed
+- Remove unimplemented security claims and retired cloud code
+- Add missing experimental feature gate to zero_trust_demo
+- Add missing experimental feature gate to zero_trust_demo
+- Resolve CodeQL alerts and README formatting
+- Resolve clippy::chunks_exact_to_as_chunks lint regression on main (#61)
+- Resolve workflow code scanning alerts (#53)
+- Give relations a distinct magic byte (0xFA), accept legacy 0xFE on read
+- Clippy needless_range_loop in resonator-bundle (§21 bin) — enumerate rewrite, outcome-neutral
+- Close remaining provenance trust-anchor gaps (cawg, vc-only records)
+- Use sort_by_key to satisfy clippy 1.96 unnecessary_sort_by
+- Repair --all-features build (port SCITT to ureq 3 API, drop dead CliffordVec bench, gate provenance example)
+- Allow CDLA-Permissive-2.0 license from webpki-roots via ureq
+- Remove dead CliffordVec code and fix clippy warnings
+- Replace bare unwrap() with expect() or total_cmp in production code
+- Perf and safety improvements from audit
+- 5 bugs from audit (deadlock, double-decay, over-count, double-start, stale-used)
+- Cargo fmt
+- Use sort_by_key for clippy on newer Rust
+- Cargo fmt formatting
+- Benchmark crate rename, cargo fmt
+- Resolve all clippy warnings for CI
+- Quality pass on G1-G4 code
+- Use absolute URL for logo on npmjs, add project.yaml for orchestrator
+- Resolve 15 audit findings (3 critical, 8 high, 4 medium)
+- Crash-safety, error propagation, and compaction correctness
+- Propagate errors from shard insert/remove, guard ShardManager invariant
+- Use .clamp() instead of .max().min() to satisfy clippy
+
+### Performance
+- Optimize Zero-Trust FHE-Lite key generation and eliminate string cloning in hot loops
+- Use posting-list candidate generation for synthesizeConcepts (>500 vectors)
+
+### Security
+- Remediate repository alerts and dependencies (#38)
+- Harden signing-key file perms, zeroize copies, bound JUMBF recursion
+
+### Assets
+- Simplify logo — strip floating decorative specks, keep the connected graph mark
+- Replace logo with self-contrasting node-graph mark (animated gif + static svg), drop garbled-text variants
+- Static transparent logo from first frame, text removed
+
+### Deps
+- Land 10 Dependabot major bumps with API fixes (#26)
+
+### Example
+- Story_memory -- WARM reasoning memory serving a writer (scrivener-mcp)
+
+### Research
+- Fractional-power encoding gives the phasor memory continuous 'near'
+- Phase memory retrieves under load where sparse collapses
+- Rotation binding gives relation algebra the current substrate lacks
+- Path-plasticity -- holographic generalization a cache can't do
+- Living-connection-graph slice -- plasticity beats saturation, verifiably
+- #2 density-matched control rules out the confound
+- #2 involution control disconfirms the self-inverse framing
+- #2 hardening RETRACTS the sparse-wins claim (AUC artifact)
+- #2 step-1 binding discriminator harness + result
+- #2 strong baselines HRR + MAP vs sparse permutation
+
+### Style
+- Apply rustfmt to experiment binaries
+