chore(release): v4.14.3 — write-gate fix, launcher-deps hardening, deps bump

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "cortex",
   "description": "Persistent memory for Claude Code — remembers across sessions automatically. Install and forget. Scientific retrieval backed by 97 published references.",
-  "version": "4.14.2",
+  "version": "4.14.3",
   "author": {
     "name": "Clement Deust",
     "email": "admin@ai-architect.tools"

CHANGELOG.md

@@ -6,6 +6,21 @@ adheres to [Semantic Versioning](https://semver.org/).
 
 ## [Unreleased]
 
+## [4.14.3] - 2026-07-17
+
+### Fixed
+- `write_class="deliberate"` (or omitted, source-fallback) was rejected by the novelty gate, violating the documented contract that deliberate writes are never rejected for low novelty. `write_class` is now threaded through `evaluate_gate` / `_compute_gate_decision` / `determine_bypass`, checked last so a more specific content-based bypass reason (`bypass_error`/`bypass_decision`/`bypass_important_tag`) still wins when it applies (#147, #148).
+- `force=True` (and plain) writes intermittently raised a misleading "check DATABASE_URL" hint on a bare `FileNotFoundError` — root cause was `validate_memory.grade_from_content(base_dir=os.getcwd())` in the write-time provenance grading step, unguarded unlike every sibling enrichment step, raising when the process cwd had been removed mid-session (e.g. a worktree cleanup), unrelated to the DB. Wrapped in the same defensive try/except pattern used elsewhere; `tool_error_handler.py`'s blanket DATABASE_URL hint no longer fires for exception types it doesn't recognize as DB-related (#147, #148).
+- Issue #149's Python-3.10-only flake: `pip_install`'s per-entry commit loop pruned superseded `*.dist-info` siblings immediately after each entry committed. `os.listdir()` order is unspecified by the stdlib and differs by OS/filesystem, so the prune could permanently delete the still-valid OLD dist-info right before the package-directory entry failed and rolled back, leaving `deps_dir` with reverted package files but no metadata for either version. The destructive prune now waits until the whole `tmp_dir` commits successfully, independent of listdir enumeration order (#149, #150).
+
+### Changed
+- `mcp` dependency bumped 1.27.0 -> 1.28.1 (uv group, dependabot) (#152).
+- `wiki_classifier.py` and `wiki_axis_registry.py` split into cohesive collaborators (`wiki_axis_defaults.py`, `wiki_classifier_gates.py`, `wiki_classifier_patterns.py`, `wiki_kind_detection.py`, `wiki_title.py`) to bring both files under the repo's 500-line limit; no behaviour change (#134, #153).
+
+### Verified
+- Pre-tag guard on the exact release tree `4e3a202b` (`benchmarks/reproduce.sh --no-ablation` + 2× `--only locomo`, isolated ephemeral pgvector containers, `reranker_state: "loaded"` in all 3 MANIFESTs, same model sha256): LongMemEval-S MRR **0.9166** (floor 0.914, +0.0026 PASS) / R@10 **0.9820** (floor 0.982, +0.0000 PASS); LoCoMo 3-run mean MRR **0.7998** (reps 0.8013 / 0.7983 / 0.7997) / R@10 **0.9135** (reps 0.9142 / 0.9127 / 0.9137, floor 0.915 tol 0.005 PASS); BEAM-100K MRR **0.5417** (not gated, inside the documented 0.539–0.547 intra-day noise band). Evidence: `benchmarks/results/repro/20260717-v4.14.3-pretag/`.
+- Adjudication (LoCoMo MRR mean 0.7998 vs threshold 0.800 = floor 0.805 − tol 0.005, i.e. −0.0002): accepted as sampling noise by explicit maintainer decision (2026-07-17). Grounds: delta vs the v4.14.2 pre-tag mean (0.8009) is −0.0011, below the standard error of a 3-rep mean (~0.0013, single-rep stdev 0.0022 per the floors-rebaseline data); rep1 under the identical full-run protocol reads 0.8013 vs 4.14.2's 0.8015; the reranker is verified loaded in every run (FlashRank-absence signature excluded); and none of the 4 released commits is in the LoCoMo harness's dependency graph — ingestion goes through `BenchmarkDB`, not the #148 write path (`benchmarks/lib/bench_db.py`). Symmetric precedent: v4.14.1 released at +0.0002 above the same threshold. #150/#152/#153 do not touch the recall path reproduce.sh's floors exercise; #148 touches the write-gate (write path) — the empirical floor-gate result above is the actual non-regression evidence, not an import-closure argument.
+
 ## [4.14.2] - 2026-07-15
 
 ### Fixed

benchmarks/results/repro/20260717-v4.14.3-pretag/MANIFEST.json

@@ -0,0 +1,25 @@
+{
+  "git_sha": "4e3a202b10c897fbc564b79d11314b2c0a1c5449",
+  "longmemeval_dataset_sha256": "08d8dad4be43ee2049a22ff5674eb86725d0ce5ff434cde2627e5e8e7e117894",
+  "pg_image": "pgvector/pgvector:pg16",
+  "bench_container_name": "cortex-bench-pg-47873-501b3391",
+  "bench_container_port": 32771,
+  "bench_runner_pid": 47873,
+  "python": "3.12.11",
+  "packages": {
+    "datasets": "4.8.4",
+    "sentence-transformers": "5.4.1",
+    "torch": "2.11.0",
+    "psycopg": "3.3.3",
+    "psycopg-pool": "3.3.0"
+  },
+  "embedding_model_revision": "1110a243fdf4706b3f48f1d95db1a4f5529b4d41",
+  "reranker_active": true,
+  "reranker_state": "loaded",
+  "reranker_model_sha256": "d3dd7b09fcf06b0c070081d6819b5effbb40b667bcad78b2d7543400271141d1",
+  "results_files": [
+    "beam-100K.json",
+    "locomo.json",
+    "longmemeval-s.json"
+  ]
+}

benchmarks/results/repro/20260717-v4.14.3-pretag/beam-100K.json

@@ -0,0 +1,69 @@
+{
+  "overall_mrr": 0.5416872795414462,
+  "overall_r10": 0.7163888888888889,
+  "ability_mrr": {
+    "abstention": 0.125,
+    "contradiction_resolution": 0.9291666666666666,
+    "event_ordering": 0.36908730158730163,
+    "information_extraction": 0.5669047619047619,
+    "instruction_following": 0.2961111111111111,
+    "knowledge_update": 0.8916666666666666,
+    "multi_session_reasoning": 0.7072916666666667,
+    "preference_following": 0.40884920634920635,
+    "summarization": 0.3019620811287478,
+    "temporal_reasoning": 0.8208333333333334
+  },
+  "ability_r5": {
+    "abstention": 0.125,
+    "contradiction_resolution": 1.0,
+    "event_ordering": 0.5,
+    "information_extraction": 0.625,
+    "instruction_following": 0.425,
+    "knowledge_update": 0.95,
+    "multi_session_reasoning": 0.9,
+    "preference_following": 0.625,
+    "summarization": 0.5,
+    "temporal_reasoning": 0.975
+  },
+  "ability_r10": {
+    "abstention": 0.125,
+    "contradiction_resolution": 1.0,
+    "event_ordering": 0.625,
+    "information_extraction": 0.65,
+    "instruction_following": 0.575,
+    "knowledge_update": 0.95,
+    "multi_session_reasoning": 0.925,
+    "preference_following": 0.7,
+    "summarization": 0.6388888888888888,
+    "temporal_reasoning": 0.975
+  },
+  "total_questions": 395,
+  "elapsed_s": 745.6345331668854,
+  "manifest": {
+    "split": "100K",
+    "n_conversations": 20,
+    "n_questions": 395,
+    "n_runs": 1,
+    "repro": {
+      "git_commit": "4e3a202b10c897fbc564b79d11314b2c0a1c5449",
+      "git_dirty": true,
+      "python_version": "3.12.11 (main, Sep 18 2025, 19:41:45) [Clang 20.1.4 ]",
+      "platform_system": "Darwin",
+      "platform_machine": "arm64",
+      "platform_node": "mac-1.home",
+      "timestamp_utc": "2026-07-17T09:51:03.497615+00:00",
+      "lib_versions": {
+        "sentence-transformers": "5.4.1",
+        "torch": "2.11.0",
+        "numpy": "2.4.4",
+        "psycopg": "3.3.3",
+        "pgvector": "0.4.2",
+        "flashrank": "0.2.10"
+      },
+      "reranker_active": true,
+      "reranker_state": "loaded",
+      "reranker_model_path": "/Users/cdeust/.cache/flashrank/ms-marco-MiniLM-L-12-v2/flashrank-MiniLM-L-12-v2_Q.onnx",
+      "reranker_model_sha256": "d3dd7b09fcf06b0c070081d6819b5effbb40b667bcad78b2d7543400271141d1"
+    }
+  }
+}

benchmarks/results/repro/20260717-v4.14.3-pretag/locomo-rep1.json

@@ -0,0 +1,52 @@
+{
+  "overall_mrr": 0.8013402367337785,
+  "overall_recall10": 0.9142280524722503,
+  "category_mrr": {
+    "multi_hop": 0.7284725312762697,
+    "temporal": 0.576656314699793,
+    "single_hop": 0.7338511764043678,
+    "open_domain": 0.8462992846762168,
+    "adversarial": 0.8580281514698554
+  },
+  "category_recall10": {
+    "multi_hop": 0.8411214953271028,
+    "temporal": 0.8043478260869565,
+    "single_hop": 0.9361702127659575,
+    "open_domain": 0.9334126040428062,
+    "adversarial": 0.9394618834080718
+  },
+  "elapsed_s": 2255.837089776993,
+  "consolidation_total_wall_s": 0.0,
+  "consolidation_call_count": 0,
+  "manifest": {
+    "with_consolidation": false,
+    "ablate_mechanism": null,
+    "ablate_env_var": null,
+    "n_conversations": 10,
+    "n_questions": 1982,
+    "n_runs": 1,
+    "consolidation_call_count": 0,
+    "consolidation_total_wall_s": 0.0,
+    "repro": {
+      "git_commit": "4e3a202b10c897fbc564b79d11314b2c0a1c5449",
+      "git_dirty": true,
+      "python_version": "3.12.11 (main, Sep 18 2025, 19:41:45) [Clang 20.1.4 ]",
+      "platform_system": "Darwin",
+      "platform_machine": "arm64",
+      "platform_node": "mac-1.home",
+      "timestamp_utc": "2026-07-17T09:13:20.977255+00:00",
+      "lib_versions": {
+        "sentence-transformers": "5.4.1",
+        "torch": "2.11.0",
+        "numpy": "2.4.4",
+        "psycopg": "3.3.3",
+        "pgvector": "0.4.2",
+        "flashrank": "0.2.10"
+      },
+      "reranker_active": true,
+      "reranker_state": "loaded",
+      "reranker_model_path": "/Users/cdeust/.cache/flashrank/ms-marco-MiniLM-L-12-v2/flashrank-MiniLM-L-12-v2_Q.onnx",
+      "reranker_model_sha256": "d3dd7b09fcf06b0c070081d6819b5effbb40b667bcad78b2d7543400271141d1"
+    }
+  }
+}

benchmarks/results/repro/20260717-v4.14.3-pretag/locomo-rep2.json

@@ -0,0 +1,52 @@
+{
+  "overall_mrr": 0.798301579293002,
+  "overall_recall10": 0.9127144298688193,
+  "category_mrr": {
+    "multi_hop": 0.7153661672353261,
+    "temporal": 0.5647299861973775,
+    "single_hop": 0.7322272880783519,
+    "open_domain": 0.8463861049770681,
+    "adversarial": 0.8572807673143996
+  },
+  "category_recall10": {
+    "multi_hop": 0.8348909657320872,
+    "temporal": 0.7934782608695652,
+    "single_hop": 0.9361702127659575,
+    "open_domain": 0.9322235434007135,
+    "adversarial": 0.9417040358744395
+  },
+  "elapsed_s": 2044.8194921016693,
+  "consolidation_total_wall_s": 0.0,
+  "consolidation_call_count": 0,
+  "manifest": {
+    "with_consolidation": false,
+    "ablate_mechanism": null,
+    "ablate_env_var": null,
+    "n_conversations": 10,
+    "n_questions": 1982,
+    "n_runs": 1,
+    "consolidation_call_count": 0,
+    "consolidation_total_wall_s": 0.0,
+    "repro": {
+      "git_commit": "4e3a202b10c897fbc564b79d11314b2c0a1c5449",
+      "git_dirty": true,
+      "python_version": "3.12.11 (main, Sep 18 2025, 19:41:45) [Clang 20.1.4 ]",
+      "platform_system": "Darwin",
+      "platform_machine": "arm64",
+      "platform_node": "mac-1.home",
+      "timestamp_utc": "2026-07-17T10:03:32.713647+00:00",
+      "lib_versions": {
+        "sentence-transformers": "5.4.1",
+        "torch": "2.11.0",
+        "numpy": "2.4.4",
+        "psycopg": "3.3.3",
+        "pgvector": "0.4.2",
+        "flashrank": "0.2.10"
+      },
+      "reranker_active": true,
+      "reranker_state": "loaded",
+      "reranker_model_path": "/Users/cdeust/.cache/flashrank/ms-marco-MiniLM-L-12-v2/flashrank-MiniLM-L-12-v2_Q.onnx",
+      "reranker_model_sha256": "d3dd7b09fcf06b0c070081d6819b5effbb40b667bcad78b2d7543400271141d1"
+    }
+  }
+}

benchmarks/results/repro/20260717-v4.14.3-pretag/locomo-rep3.json

@@ -0,0 +1,52 @@
+{
+  "overall_mrr": 0.7997160956545416,
+  "overall_recall10": 0.91372351160444,
+  "category_mrr": {
+    "multi_hop": 0.7229528259902092,
+    "temporal": 0.5720194962042788,
+    "single_hop": 0.7358057525610717,
+    "open_domain": 0.8450611516901647,
+    "adversarial": 0.8568385650224215
+  },
+  "category_recall10": {
+    "multi_hop": 0.8348909657320872,
+    "temporal": 0.8152173913043478,
+    "single_hop": 0.9397163120567376,
+    "open_domain": 0.9310344827586207,
+    "adversarial": 0.9417040358744395
+  },
+  "elapsed_s": 2060.7755172252655,
+  "consolidation_total_wall_s": 0.0,
+  "consolidation_call_count": 0,
+  "manifest": {
+    "with_consolidation": false,
+    "ablate_mechanism": null,
+    "ablate_env_var": null,
+    "n_conversations": 10,
+    "n_questions": 1982,
+    "n_runs": 1,
+    "consolidation_call_count": 0,
+    "consolidation_total_wall_s": 0.0,
+    "repro": {
+      "git_commit": "4e3a202b10c897fbc564b79d11314b2c0a1c5449",
+      "git_dirty": true,
+      "python_version": "3.12.11 (main, Sep 18 2025, 19:41:45) [Clang 20.1.4 ]",
+      "platform_system": "Darwin",
+      "platform_machine": "arm64",
+      "platform_node": "mac-1.home",
+      "timestamp_utc": "2026-07-17T10:37:43.471538+00:00",
+      "lib_versions": {
+        "sentence-transformers": "5.4.1",
+        "torch": "2.11.0",
+        "numpy": "2.4.4",
+        "psycopg": "3.3.3",
+        "pgvector": "0.4.2",
+        "flashrank": "0.2.10"
+      },
+      "reranker_active": true,
+      "reranker_state": "loaded",
+      "reranker_model_path": "/Users/cdeust/.cache/flashrank/ms-marco-MiniLM-L-12-v2/flashrank-MiniLM-L-12-v2_Q.onnx",
+      "reranker_model_sha256": "d3dd7b09fcf06b0c070081d6819b5effbb40b667bcad78b2d7543400271141d1"
+    }
+  }
+}

benchmarks/results/repro/20260717-v4.14.3-pretag/longmemeval-s.json

@@ -0,0 +1,54 @@
+{
+  "overall_mrr": 0.9166380952380953,
+  "overall_recall10": 0.982,
+  "category_mrr": {
+    "Single-session (user)": 0.8412698412698413,
+    "Multi-session reasoning": 0.9625313283208021,
+    "Single-session (preference)": 0.6853703703703703,
+    "Temporal reasoning": 0.9184389545291801,
+    "Knowledge updates": 0.9320512820512821,
+    "Single-session (assistant)": 1.0
+  },
+  "category_recall10": {
+    "Single-session (user)": 0.9571428571428572,
+    "Multi-session reasoning": 1.0,
+    "Single-session (preference)": 0.9,
+    "Temporal reasoning": 0.9774436090225563,
+    "Knowledge updates": 1.0,
+    "Single-session (assistant)": 1.0
+  },
+  "elapsed_s": 1987.0532216249994,
+  "consolidation_total_wall_s": 0.0,
+  "consolidation_call_count": 0,
+  "manifest": {
+    "with_consolidation": false,
+    "with_consolidation_note": "Scores collected with consolidation=False do NOT reflect production behaviour.  Consolidation-only mechanisms (CASCADE, INTERFERENCE, HOMEOSTATIC_PLASTICITY, SYNAPTIC_PLASTICITY, MICROGLIAL_PRUNING, TWO_STAGE_MODEL, EMOTIONAL_DECAY, TRIPARTITE_SYNAPSE, SCHEMA_ENGINE) are exercised only when with_consolidation=True.  Delta between the two conditions is unmeasured in this run.",
+    "ablate_mechanism": null,
+    "ablate_env_var": null,
+    "n_questions": 500,
+    "n_runs": 1,
+    "consolidation_call_count": 0,
+    "consolidation_total_wall_s": 0.0,
+    "repro": {
+      "git_commit": "4e3a202b10c897fbc564b79d11314b2c0a1c5449",
+      "git_dirty": true,
+      "python_version": "3.12.11 (main, Sep 18 2025, 19:41:45) [Clang 20.1.4 ]",
+      "platform_system": "Darwin",
+      "platform_machine": "arm64",
+      "platform_node": "mac-1.home",
+      "timestamp_utc": "2026-07-17T08:40:12.173333+00:00",
+      "lib_versions": {
+        "sentence-transformers": "5.4.1",
+        "torch": "2.11.0",
+        "numpy": "2.4.4",
+        "psycopg": "3.3.3",
+        "pgvector": "0.4.2",
+        "flashrank": "0.2.10"
+      },
+      "reranker_active": true,
+      "reranker_state": "loaded",
+      "reranker_model_path": "/Users/cdeust/.cache/flashrank/ms-marco-MiniLM-L-12-v2/flashrank-MiniLM-L-12-v2_Q.onnx",
+      "reranker_model_sha256": "d3dd7b09fcf06b0c070081d6819b5effbb40b667bcad78b2d7543400271141d1"
+    }
+  }
+}

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "hypermnesia-mcp"
-version = "4.14.2"
+version = "4.14.3"
 description = "Scientifically-grounded memory system based on computational neuroscience research"
 readme = "README.md"
 license = "MIT"