Define a scaffold compatibility and rollout manifest for trainable RLM roots

[shanev]

Why

Post-trained RLM roots learn against a scaffold, not against an abstract "RLM" interface. Prompt templates, REPL globals, capability names, terminal semantics, truncation, errors, and budgets jointly form the model-facing ABI. A rollout is not reproducible or safe to use for training unless it records that resolved contract.

Droste already has versioned runner/source protocols, prompt-pack identity, and a unified Trace ABI. The missing public contract is one immutable, content-addressed manifest that identifies the exact scaffold and inference configuration that produced a rollout or that a checkpoint expects.

Scope

• Define a strict, versioned ScaffoldManifest / RolloutManifest schema containing at least:
  • Droste engine version and source revision when available;
  • runner, source, kernel/capability, trace, and prompt-pack ABI versions;
  • resolved prompt-pack ID, revision, profile, and content hash;
  • model-visible globals/capability manifest and its hash;
  • terminal, output-truncation, refinement, repair, and extraction contract identities;
  • root and subcall model IDs/immutable revisions;
  • sampling parameters, output limits, budgets, timeout, concurrency, and seed;
  • parent/child identity semantics for subcalls.
• Produce the manifest once at run start after all defaults and overrides resolve.
• Include its ID/hash in startup, result/replay, and terminal run evidence without copying sensitive prompt or data content into durable telemetry.
• Add a compatibility check that can compare a checkpoint's declared scaffold requirements with a resolved run and return a typed mismatch before inference.
• Define an extension/join seam for trainer-owned task, split, verifier, reward, and authorization metadata. Keep those evaluation facts outside Droste's generic runtime trace.
• Document canonical serialization, hashing, optional-field rules, and compatibility behavior.

Design constraints

• Do not couple Droste to a trainer, model registry, database, or product.
• Do not treat trace retention as permission to train.
• Do not make prompt text, user context, code, stdout, or answers part of content-free durable telemetry.
• Do not duplicate PromptPack, capability, Trace ABI, or runner-protocol types; compose their identities.
• Adding a trainer adapter must not change the engine's pure runtime contract.

Tests

• Equivalent resolved runs serialize byte-identically and yield the same hash.
• A material prompt-pack, capability, terminal-contract, model-revision, sampling, or budget change changes the hash.
• Dictionary ordering and irrelevant host metadata do not change the hash.
• Missing/incompatible checkpoint requirements fail with a typed, actionable error before an LLM request.
• Durable evidence remains content-free; replay remains retention-controlled.
• Native and runner/Pyodide paths resolve the same manifest for the same effective configuration.

Acceptance criteria

• A public versioned manifest schema and canonical hash algorithm are documented.
• Every Droste run exposes the resolved manifest identity.
• Checkpoint compatibility can be checked without importing a trainer or model registry.
• Trainer-owned outcome metadata has an explicit external join seam.
• Trace retention and training authorization remain independent.
• Tests cover determinism, compatibility failures, privacy, and native/runner parity.

Related public work

• Prompt packs + RLM skills: move all strategy out of the engine into versioned data #3 — versioned prompt packs and RLM skills
• Unified Trace ABI: one versioned run record for trajectory, events, usage, policy, capability calls #13 — unified Trace ABI

[shanev]

Implementation update: PR #112 implements the public scaffold-manifest value, lossless content-addressed identity, runner/native parity, checkpoint requirement comparison before the first root-model call, and content-free terminal provenance.

The review removed an initially proposed stdout truncation bit because the engine rejects oversized output and has no authoritative truncation signal; stdout character counts remain exact. ModelRelay serving integration is tracked in tensor-systems/modelrelay#1728 and requires a post-release vendored Droste bump. This issue should remain open until that end-to-end serving path and rollout evidence are complete.

[shanev]

Host-integration update: PR #117 adds runner protocol v4 scaffold preflight. It uses the exact same public resolver as real execution, returns the complete content-free manifest, preserves typed ScaffoldCompatibilityError mismatch fields, and makes no root/subcall/provider calls or live provider bindings.

The v4 bump is a safety boundary: older v3 runners reject the request before work instead of ignoring an unknown preflight operation and executing. This issue remains open until the released tag is synced and the hosted end-to-end refusal is proven.