feat(improvement): train and serve a checkpoint through improve with a receipt

CHANGELOG.md

@@ -1,5 +1,13 @@
 # Changelog
 
+## 0.242.0
+
+`improve(profile, { mode: 'training', ... })` and the bound harness's `train` method now produce a checkpoint-backed candidate with an Interface training receipt. The command trainer pins its executable and inputs, supplies only an explicit public environment, and cancels its POSIX process group. Managed trainers and verified serving adapters use the same typed boundary; they own remote job cleanup.
+
+The dataset envelope preserves existing Eval row payloads, inventories every exposed train/validation task, and refuses cross-partition reuse. Runtime verifies checkpoint bytes, artifact-addressed serving identity, and complete receipt ancestry. It reuses the existing cancellation and durable-file primitives: pre-dispatch cancellation starts no adapter, candidate validation cannot silently change checkpoint bytes, and the receipt is durable before profile publication. A late cancellation does not retract a committed profile; failure paths report uncertainty about external jobs.
+
+Requires `@tangle-network/agent-interface` `^2.10.0`, including the exact packed cohort. Training constructs a candidate; evaluation and promotion remain with the existing held-out gates. No GPU scheduler, managed trainer implementation, or Router deployment adapter is added.
+
 ## 0.241.1
 
 Coordination public-address signals now end when their listener closes or setup fails, as well as on manager cancellation. Hosted ingress adapters can retire a route before its local port is reused, without changing authentication, execution policy or agent capabilities.

api-surface.json

@@ -29,13 +29,15 @@
       "CandidateGenerator": "type d2b99fd9843d",
       "ChatModelCandidate": "type c35e223e502c",
       "ChatModelValidation": "type fc44fbff36e7",
+      "CheckpointServingPort": "type 7bf8caed3052",
       "CircuitBreakerConfig": "type d7631a064800",
       "ConfigError": "value a04b1851d315",
       "ControlBudget": "type b0404d2cb0b4",
       "ControlDecision": "type 251bcb710b26",
       "ControlEvalResult": "type c7a2d96ce0ac",
       "ControlRunResult": "type 4994c45c1afc",
       "ControlStep": "type 69d8ab196483",
+      "ControlledTrainingCommand": "type e787701331f6",
       "Conversation": "type da37c7083ab5",
       "ConversationDriveState": "type 7320e9a7f523",
       "ConversationJournal": "type a8842d8d1d54",
@@ -92,6 +94,8 @@
       "ImproveScenarioPartitions": "type 37a3508406b1",
       "ImproveSkillsOptions": "type c1f5a69faefc",
       "ImproveSurface": "type b711b683b151",
+      "ImproveTrainingOptions": "type 8bfc42270e58",
+      "ImproveTrainingResult": "type 81c28cf9d5ab",
       "ImprovementCandidate": "type 0c22a91c6396",
       "ImprovementCodeCandidate": "type 588fa6d3b2f5",
       "ImprovementMaterializedProfilePopulationCandidate": "type 4eee039de8cf",
@@ -142,8 +146,11 @@
       "PersonaConversationResult": "type d642750715ee",
       "PersonaDriver": "type 14b302d4c33f",
       "PlannerError": "value 3b3330bf4164",
-      "ProfileImprovementHarness": "type aa5280c89d45",
+      "ProfileImprovementHarness": "type cde3e185f5f2",
       "ProfileImprovementHarnessRunOptions": "type 3d1ff1d78855",
+      "ProfileImprovementHarnessTrainOptions": "type f1982d2b20f7",
+      "ProfileTrainer": "type 2907ac3022c3",
+      "ProfileTrainerRequest": "type 542a10bc9b12",
       "RESEARCH_SUPERVISOR_SYSTEM_PROMPT": "value 373728f5643d",
       "ROLLOUT_POLICY_EXTENSION": "value 4fa0cbedb455",
       "RawTraceDistillerOptions": "type a04b8b0d09e1",
@@ -203,6 +210,8 @@
       "SupervisedKnowledgeUpdateOptions": "type e20062b4394e",
       "SupervisedKnowledgeUpdateResult": "type ac482862a4ae",
       "SupervisedKnowledgeUpdater": "type 90a6812eba78",
+      "TrainingBoundaryResult": "type f977a49cfd72",
+      "TrainingDatasetDocument": "type 5a17f19e9620",
       "TurnOrder": "type 573c9a049337",
       "ValidationError": "value 44437c30eb82",
       "Verifier": "type c596e7edac69",
@@ -222,6 +231,7 @@
       "commandVerifier": "value 75c0a10aca5d",
       "composeRuntimeHooks": "value a0c8fbe0631b",
       "createAgentKnowledgeReadinessCheck": "value 86edaf4f4ca1",
+      "createCommandProfileTrainer": "value 36c8ccf5d298",
       "createConversationBackend": "value 9fe923adfb90",
       "createIterableBackend": "value 314afcb51e55",
       "createKnowledgeImprovementActivationExecutor": "value b6efdb7cc158",
@@ -243,7 +253,7 @@
       "formatSupervisedKnowledgeTask": "value bdcf6b28157d",
       "generateSpanId": "value 2f8329045cac",
       "getModels": "value 95cb7c012c48",
-      "improve": "value 1303859a2197",
+      "improve": "value 45f11e9feb52",
       "isDelegatedLoopMode": "value d0f2042750ec",
       "knowledgeReadinessDeliverable": "value f6f33b24a926",
       "loopEventToOtelSpan": "value 63bec8b09ae0",

bench/CHANGELOG.md

@@ -1,5 +1,10 @@
 # Changelog
 
+## 0.13.4
+
+Require Interface `^2.10.0` and consume Runtime 0.242.0 through the published dependency ranges, keeping benchmark consumers on the checkpoint-training receipt contract.
+Benchmark execution and grading behavior are unchanged.
+
 ## 0.13.3
 
 Support Sandbox 0.41 and consume Runtime 0.233.1 through the published dependency ranges.

bench/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-bench",
-  "version": "0.13.3",
+  "version": "0.13.4",
   "type": "module",
   "description": "Benchmark adapters and execution for agent-runtime across coding, tool-use, RAG, memory, browser, and terminal tasks.",
   "repository": {