@a9i5k4/dsh-auto-memory 3.1.5 → 3.1.6

package/python/m7_activation_features_v2.py

@@ -1,11 +1,11 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
-"""M7 activation features pre v2 (policy: activation_policy_v2).
+"""M7 activation features pre v2 (policy: activation_policy_pre_v2).
 
 Round-1 implementation of the two-lane activation eligibility decision.
 Pure stdlib; deterministic; NO pickle/joblib/sklearn at runtime — the
 calibrated intent head is reproduced from the auditable JSON artifact
-python/policies/recall_intent_lr_v1.json (vocabulary + IDF + logistic
+python/policies/recall_intent_lr_pre_v1.json (vocabulary + IDF + logistic
 regression coefficients/intercept + Platt calibration), exported by the
 calibration pipeline (runId label-review-cal20260824-1954).
 
@@ -48,9 +48,9 @@ import math
 import os
 import re
 
-FEATURES_POLICY_VERSION = 'activation_features_v2'
-INTENT_POLICY_VERSION = 'recall_intent_lr_v1'
-ACTIVATION_POLICY_VERSION = 'activation_policy_v2'
+FEATURES_POLICY_VERSION = 'activation_features_pre_v2'
+INTENT_POLICY_VERSION = 'recall_intent_lr_pre_v1'
+ACTIVATION_POLICY_VERSION = 'activation_policy_pre_v2'
 
 INTERROG = ['什么', '如何', '怎么', '哪些', '哪个', '为什么', '多少', '吗',
             '呢', '啥', 'recall', 'what', 'how', 'which', 'when', 'where',
@@ -375,8 +375,8 @@ if __name__ == '__main__':
     here = os.path.dirname(os.path.abspath(__file__))
     pol_dir = os.path.join(here, 'policies')
     loaded = load_and_verify_policy(
-        os.path.join(pol_dir, 'recall_intent_lr_v1.json'),
-        os.path.join(pol_dir, 'activation_policy_v2.json'))
+        os.path.join(pol_dir, 'recall_intent_lr_pre_v1.json'),
+        os.path.join(pol_dir, 'activation_policy_pre_v2.json'))
     demo_rows = [
         {'id': 'demo-recall', 'text': '之前为什么选 BGE-M3？',
          'denseTop': 0.72, 'margin': 0.10, 'containment': 0.55, 'mark': 1,

package/python/m7_embedding_v1.py

@@ -6,17 +6,17 @@ Pure re-implementation of the frozen policy for the production sidecar path
 (the benchmark rig lives separately under python/bench/ and stays untouched):
 
   provider      bge-m3 pinned revision, CLS pooling, L2-normalized float32
-  chunk policy  m7_chunk_v1 = para-512-noov (tokenizer id space,
+  chunk policy  m7_chunk_pre_v1 = para-512-noov (tokenizer id space,
                 greedy paragraph packing, oversized paragraph hard-split,
                 no overlap; special tokens added by the model tokenizer)
   identity      provider/model/revision/dimension/normalization/policyVersion
                 -> configHash; mismatch = stale = full rebuild
 
 Providers:
-  hash-v1   deterministic stdlib-only embedding (sha256-seeded bag of
+  hash-pre-v1   deterministic stdlib-only embedding (sha256-seeded bag of
                 token-trigram dims). Zero dependencies, zero network. Used
                 by CI and offline protocol tests. NOT a quality provider.
-  bge-m3-v1 real model via transformers (lazy import; requires the
+  bge-m3-pre-v1 real model via transformers (lazy import; requires the
                 pinned local snapshot dir passed in the embedding config).
 
 No DSH file reads; no writes except what the worker explicitly passes in.
@@ -25,15 +25,15 @@ import hashlib
 import json
 import re
 
-PROVIDER_REAL = 'bge-m3-v1'
-PROVIDER_REAL_INT8 = 'bge-m3-onnx-int8-v1'
-PROVIDER_HASH = 'hash-v1'
-CHUNK_POLICY_VERSION = 'm7_chunk_v1'
+PROVIDER_REAL = 'bge-m3-pre-v1'
+PROVIDER_REAL_INT8 = 'bge-m3-onnx-int8-pre-v1'
+PROVIDER_HASH = 'hash-pre-v1'
+CHUNK_POLICY_VERSION = 'm7_chunk_pre_v1'
 CHUNK_MAX_TOKENS = 512
 QUERY_MAX_TOKENS = 256
 DIMENSION = 1024
 
-RE_CHUNK = re.compile(r'^chk_[0-9a-f]{16,}$')
+RE_CHUNK = re.compile(r'^chk_pre_[0-9a-f]{16,}$')
 
 
 def sha_hex(data):
@@ -73,8 +73,8 @@ def config_hash(provider, model_revision, dimension):
 
 
 def chunk_id_for(memory_id, record_digest, ordinal):
-    return 'chk_' + sha_hex(
-        ('m7-chunk-v1\u0000' + memory_id + '\u0000' + record_digest +
+    return 'chk_pre_' + sha_hex(
+        ('m7-chunk-pre-v1\u0000' + memory_id + '\u0000' + record_digest +
          '\u0000' + str(ordinal)).encode('utf-8'))[:32]
 
 
@@ -359,8 +359,8 @@ def identity_block(provider, config):
         dtype = 'int8-dynamic-onnx'
     return {
         'schemaVersion': 1,
-        'namespace': 'dsh-auto-memory',
-        'policyVersion': 'semantic_vectors_v1',
+        'namespace': 'dsh-auto-memory-pre',
+        'policyVersion': 'semantic_vectors_pre_v1',
         'provider': provider,
         'model': model_name,
         'modelRevision': str(config.get('modelRevision') or 'hash'),

package/python/policies/activation_policy_v2.json

@@ -1,7 +1,7 @@
 {
  "schemaVersion": 1,
- "policyVersion": "activation_policy_v2",
- "parentPolicyVersion": "m7_semantic_threshold_v1",
+ "policyVersion": "activation_policy_pre_v2",
+ "parentPolicyVersion": "m7_semantic_threshold_pre_v1",
  "createdAt": "2026-08-25",
  "runId": "label-review-cal20260824-1954",
  "goldDigest": "e74ccdcda998ae51596adf8119c818ac84fc52a4fa657fd687daeda7d14a072d",

package/python/policies/decision-record-activation-v2-delta-exp-override-20260824.json

@@ -1,7 +1,7 @@
 {
  "schemaVersion": 1,
  "decisionId": "activation-v2-delta-exp-override-20260824",
- "policyVersion": "activation_policy_v2",
+ "policyVersion": "activation_policy_pre_v2",
  "previousDeltaExp": 0.02,
  "approvedDeltaExp": 0.03,
  "reason": "production completeness semantics exposed cal-0008 prefetch-to-emit violation at 0.02",

package/python/policies/decision-record-reasoning-kind-admission-20260826.json

@@ -7,8 +7,8 @@
   "summary": "'reasoning' admitted as a legal Segment kind (weight 0.5 in lexical query plan), superseding the M4/M5-era rule that reasoning never enters context.",
   "trigger": "controlled live shadow round 2: envelope drops 'trigger:invalid:kind' + 'window:invalid:kind' — the new CoT segments (kind='reasoning') were rejected by the M5 whitelist ['user','tool_call','tool_result','assistant']",
   "change": [
-    "lib/context-bridge.js: segment kind whitelist + 'reasoning'",
-    "lib/shadow-retrieval.js: trigger.kind whitelist + 'reasoning'; query-plan origin map adds 'reasoning' at weight 0.5 (between recent-user 0.8 and tool-result 0.6... actually below tool-result, above assistant 0.2)",
+    "lib/context-bridge-pre.js: segment kind whitelist + 'reasoning'",
+    "lib/shadow-retrieval-pre.js: trigger.kind whitelist + 'reasoning'; query-plan origin map adds 'reasoning' at weight 0.5 (between recent-user 0.8 and tool-result 0.6... actually below tool-result, above assistant 0.2)",
     "smoke-test-m51-pre.mjs: old rejection assertion replaced with admission assertion"
   ],
   "safety": [

package/python/policies/decision-record-stale-gate-per-candidate-20260825.json

@@ -11,10 +11,10 @@
     "rootCause": "worker constructed hardGates.stale as any(e.freshness=='stale' for e in payload.evidence); on real traffic every context_push carries >=1 stale evidence entry (anchored files updated today -> all M5 historical evidence judged stale), so fv2 was permanently muted"
   },
   "change": {
-    "file": "python/worker_semantic_v1.py (_fv2_shadow_decide gate construction only)",
+    "file": "python/worker_semantic_pre_v1.py (_fv2_shadow_decide gate construction only)",
     "semantics": "stale/correction gates fire ONLY when the affected evidence memoryId is among the CURRENT top-K candidates; unrelated evidence never gates",
     "untouched": [
-      "m7_activation_features_v2.py (decision core unchanged)",
+      "m7_activation_features_pre_v2.py (decision core unchanged)",
       "python/policies/*.json (versions, configHashes unchanged: activation cfgh_31e6d977..., intent cfgh_69465e69...)",
       "golden-parity-fixtures-v1.jsonl (fixtures carry explicit hardGates inputs; all False -> zero output drift)"
     ]