[RAG Retrieval] Implement semantic and deterministic hybrid retrieval

[rica-v3]

Parent and project

• Parent: [RAG Runtime] Deliver a dependency-free generic RAG package #1
• Project: https://github.com/orgs/Kims-DeveloperGroup/projects/1
• Milestone: Generic RAG v1
• Iteration: Iteration 09
• Wave: 12
• Prerequisites completed: [RAG API] Bootstrap the package and define generic contracts #2 and [RAG Projection] Add revision-aware derived indexing and rebuilds #3
• Owner: @rica-v3

Objective

Implement bounded semantic retrieval and deterministic rank-based hybrid fusion over injected vector and lexical collaborators, then expose resource-free runtime composition suitable for independent consumers.

In scope

• Validated bounded semantic queries and opaque ranked hits.
• Deterministic reciprocal-rank-style fusion of lexical and semantic candidates without comparing vendor raw scores.
• Stable deduplication by corpus/document/revision/fragment identity and deterministic tie-breaking.
• Explicit truncation, unavailable, stale, and collaborator-failure results.
• Resource-free composition that borrows injected collaborators and never dynamically loads or closes them.
• Complete API, architecture, security-boundary, integration, and operational documentation.
• Independent package smoke and clean-wheel verification.

Out of scope

Application authorization/citations, automatic transcript expansion, vendor adapters, credentials, networking, SQLite/MySQL, dynamic imports, background indexing, or an installed CLI.

Acceptance criteria

• Search work and returned candidates are strictly bounded.
• Fusion and tie ordering are deterministic for equal inputs.
• Vendor score scale cannot affect cross-retriever ordering.
• Duplicate/stale candidates are handled through exact generic identity and revision contracts.
• Collaborator failure cannot fabricate successful or authoritative content.
• The complete package works without network, credentials, or third-party dependencies.
• Public docs and module index match the released API.

Verification

Semantic/fusion ranking matrices, deduplication/tie/budget tests, collaborator failures, clean wheel and import tests, dependency scan, full suite, compileall, docs/index review.

Dependencies and risks

Blocked by #2 contracts and #3 projection lifecycle. Risks include nondeterministic float ordering, unbounded candidate work, accidental authority claims, and coupling runtime composition to concrete providers.

Gates

github_planning_gate = pending until all native relationships, Project metadata, milestone, and branch evidence are verified. modularity_gate = required; final review must prove the generic package remains application-independent.