Prompt packs + RLM skills: move all strategy out of the engine into versioned data

[shanev]

Problem

Human strategy was historically embedded in engine constants and loop templates. Strategy must be removable, versioned data so consumers can change it without forking the engine or changing capability, budget, trace, and result contracts.

Prompt packs

• A complete versioned artifact owns the base, refinement, repair, extraction, and policy-default templates.
• The stable slot contract is {capabilities}, {budget}, {question}, {history}, and {output_contract}.
• One pack resolves at run start by (model family, profile) through caller, consumer, engine-family, and generic catalog tiers.
• Resolution is exclusive; partial packs are not merged.
• Provenance may cite immutable, version-matched artifacts. Without those artifacts it makes no effectiveness claim.

RLM skills

• Additive Markdown-plus-frontmatter artifacts teach optional code-writing strategies under a metered budget.
• Compatibility guidance can migrate from constants into versioned skills.
• Skills may be loaded through brokered capabilities and consume the same run budget.
• Consumer-owned catalogs use the public pack, skill, and provider contracts rather than parallel extension systems.

Exit path

A capabilities-and-contracts-only profile must run by swapping versioned data, without a loop change. Compatibility profiles may retain current guidance until evidence supports shrinking them.

Remaining work

Prompt packs and provenance have landed. Complete the removable-skills path and its acceptance tests while preserving intent-blind policy hints, generated capability descriptions, and one accounting/dispatch path.

[shanev]

Prompt packs and provenance have landed. The remaining work in this issue is the removable-skills half and its exit-ramp acceptance tests:

• generated capabilities and budget/output contracts remain engine-owned facts;
• broad chunking, batching, map/reduce, and semantic-workflow guidance is additive versioned data, not loop logic;
• built-in full/minimal guidance is compatibility data, while a none/capabilities-only configuration must run without changing the loop;
• consumer-owned packs/skills can replace the defaults through the public catalog/broker contracts;
• no semantic/count intent is inferred from question text; callers continue to supply explicit policy hints when they require enforcement;
• swapping or removing strategy preserves the same capability, budget, trace, and result contracts.

The public hygiene cleanup for the surrounding design prose is tracked separately in #99. No new generic strategy issue is needed.

[shanev]

Training compatibility follow-up: #101 composes PromptPack identity into a content-addressed scaffold manifest. This issue remains the owner of PromptPack loading and slot semantics.

[shanev]

Additional acceptance criteria: model-family-specific PromptPack revisions must resolve without changing the slot ABI; decomposition examples are versioned assets with stable IDs and hashes; provenance records the exact selected asset set; model-specific variants cannot silently become universal defaults; and full, minimal, and capabilities-and-contracts-only variants remain selectable by swapping versioned data.