V37 Gate 11: Preserve V38 inference scope

Expand the V38 notes-only opening with source-anchored requirements for FailsafeGenerationSequence over ThricifiedGeneration, prompt and tool composition, prompt benchmarking, telemetry, and ReadFitsFindingSynthesis depository search.

Author: geraldarthurdavis

BITCODE_SPEC_V38_NOTES.md

@@ -18,30 +18,113 @@ routes, tests, or package APIs that claim V38 implementation before Gate 1.
 
 ## Deferred from V37
 
-V37 promoted Website Conversations. V38 should inspect the inference execution
-stack beneath PTRR agents and ensure practical FailsafeGeneration layers are
-specified above ThricifiedGeneration calls where agents need context
-preparation, large input handling, large output handling, and repairable
-typed-output generation.
-
-Thricification remains the lowest-level inference call chain. It resolves final
-prompt registry composition, interpolation, and typed output. Failsafe chains
-may orchestrate setup and repair around those calls, but must delegate final
-reason, judge-reasoning, and typed response production to ThricifiedGeneration
-when used inside Reading pipeline agents.
+V37 promoted Website Conversations. V38 is inference-focused: it should correct
+remaining gaps in the pipeline inference stack, fully use Bitcode's existing
+agent, prompt, execution, tool, telemetry, and registry primitives, and bring
+all Reading inference surfaces to commercial-readiness without weakening any
+runtime boundary.
+
+Source anchors already exist and must be audited before V38 implementation:
+
+- `packages/agent-generics/src/steps/failsafe-sequence.ts` defines
+  `FailsafeGenerationSequence` as prepare-concise-context, chunk-then-sum, and
+  stitch-until-complete wrappers over `ThricifiedGeneration`.
+- `packages/agent-generics/src/steps/thricified-generation.ts` defines
+  `ThricifiedGeneration` as the strict Reason, Judge, StructuredOutput chain.
+- `packages/agent-generics/src/agents/factories.ts` requires
+  `factoryAgentWithPTRR` callers to provide an agent prompt plus complete
+  Plan, Try, Refine, and Retry step prompt registries.
+- `packages/tools-generics/src/Tool.ts`,
+  `packages/tools-generics/src/doc-code-tool/DocCodeToolPrompt.ts`, and
+  `packages/tools-generics/src/doc-code-tool/formatUsableTools.ts` define the
+  doc-comment-backed tool prompt path that inference prompts must consume.
+- `packages/pipelines/asset-pack/src/reading-pipeline-contract.ts`,
+  `packages/pipelines/asset-pack/src/read-need.ts`,
+  `packages/pipelines/asset-pack/src/depository-search.ts`, and
+  `packages/pipelines/asset-pack/src/tools/*DepositorySearchTool*` define the
+  current `ReadNeedComprehensionSynthesis` and
+  `ReadFitsFindingSynthesis` contracts, tools, search result shape, and
+  source-safe preview boundary.
+
+Thricification remains the lowest-level inference call chain. It is the final
+place where phase, agent, step, substep, tool-documentation, and generation
+prompt registries are resolved, composed, interpolated, and constrained to the
+typed output. Failsafe chains may prepare context, chunk large inputs, summarize
+or stitch large outputs, retry repairable typed-output failures, and organize
+execution ancestry state around those calls, but inside PTRR agents they must
+delegate final reason, judge-reasoning, and structured typed response
+production to `ThricifiedGeneration`.
+
+Tools remain step-owned capability executions, not hidden children of
+`ThricifiedGeneration`. Tool definitions must still be represented in final
+prompts through the doc-comment `DocCodeToolPrompt` path, composed and
+interpolated at the lowest inference point with the same source-safe telemetry
+discipline as all other prompt material.
 
 ## Candidate V38 workstreams
 
-- Audit PTRR agent step execution so phase, agent, step, tool, Failsafe, and
-  ThricifiedGeneration responsibilities are named without overlap.
-- Specify prompt registry composition from phase prompts through agent prompts,
-  step prompts, substep prompts, and final generation prompts.
+- Audit every inference point in Reading pipelines, Website Conversations,
+  tool-definition prompts, and interface-specific inference entrypoints. The
+  audit must count base primitives and specific implementations separately:
+  phase prompts, PTRR agent prompts, step prompts, Failsafe substep prompts,
+  ThricifiedGeneration prompts, tool prompt definitions, prompt templates,
+  interpolation bindings, expected context fields, raw outputs, parsed typed
+  outputs, and repair/failure surfaces.
+- Specify the exact execution stack for pipeline agents without overlapping
+  names: PipelineExecution owns phases; PTRR agents own Plan, Try, Refine, and
+  Retry steps; FailsafeGenerationSequence owns context preparation, large input
+  handling, large output handling, and repair orchestration; ThricifiedGeneration
+  owns Reason, Judge, and StructuredOutput inference calls; ToolExecution owns
+  callable tools and tool input/output receipts.
+- Require all PTRR agent steps to use practical FailsafeGenerationSequence
+  layers above ThricifiedGeneration unless a source-backed exception is
+  specified, tested, and justified. The default Reading stack should be
+  PipelineExecution -> PTRR agent -> PTRR step -> FailsafeGenerationSequence ->
+  ThricifiedGeneration -> provider call.
+- Finish prompt and PromptPart benchmarking so semantically divided
+  PromptParts and complete Prompts are both benchmarkable. V38 Gate 1 should
+  inventory existing `benchmarks:` doc-comment metadata, the
+  `packages/prompts/src/benchmarking/*` runner, and all Reading and
+  conversation prompt surfaces, then define initial benchmark suites for every
+  prompt part and prompt used by active inference.
+- Complete prompt registry composition and interpolation rules from phase
+  prompts through agent prompts, step prompts, substep prompts, final generation
+  prompts, and tool documentation. The final inference call must expose the
+  unrendered template identity, interpolated prompt, context bindings, selected
+  tools, raw response, parsed output, and schema verdict to telemetry at the
+  proper disclosure tier.
+- Integrate this inference work with prior telemetry. Execution ancestry
+  state-stack data must be written, read, prepared by Failsafe layers,
+  interpolated into prompts, maintained through retries, and streamed as
+  source-safe phase, agent, step, Failsafe, ThricifiedGeneration, tool, input,
+  output, and typed-result events.
+- Complete `ReadFitsFindingSynthesis` depository-search internals as the
+  coupled second focus of V38. Fit-finding queries are inference-derived and
+  must search throughout the Bitcode Depository using lexical, symbolic, path,
+  metadata, measurement, embedding/vector, and provider-specific channels where
+  appropriate.
+- Harden the depository-search tool and primitive utility stack around source
+  embeddings, measurement embeddings, metadata embeddings, query synthesis,
+  ranking, verification, and AssetPack synthesis handoff. Existing OpenAI
+  embedding policy (`text-embedding-3-small`, 1536 dimensions, cosine
+  `match_deliverable_vectors`) is an active source anchor, and V38 may specify
+  additional LlamaIndex, Pinecone, or provider-backed channels only when their
+  boundaries, storage, credentials, telemetry, and tests are explicit.
+- Preserve the economic carry-through from found fits to AssetPack settlement:
+  selected fit deposits must remain traceable into source-safe preview,
+  post-settlement delivery, contributor compensation, exact ledger/database
+  synchronization, and proof receipts.
 - Clarify how Bitcoin, GitHub, compute, storage, and build/process boundaries
   are represented in long-running inference and pipeline telemetry.
 - Confirm that source-safe telemetry exposes prompt templates, interpolated
   prompts, raw provider responses, parsed typed results, repair attempts, and
   failed-output evidence only at their permitted disclosure tier.
 
+The V38 acceptance posture is commercial inference correctness: Bitcode's
+ability to maximize inference intelligence must produce reliable technical
+intelligence measurement, labeling, fit-finding, AssetPack synthesis, and
+mutually beneficial depositing and Reading exchange.
+
 ## Non-goals during V38 opening
 
 - No new production route, API version, source package name, or UI label should