reviewer evidence: provide canonical trusted research retrieval handoff for Noema receipts

[seonghobae]

Problem

ContextualWisdomLab/noema#555 owns reviewer evidence admission and protected Noema source already has the generic research receipt kernel. Draft noema#607 adds only the Noema-side producer-specific byte-integrity adapter: already-retrieved immutable bytes, a non-empty exact excerpt present verbatim in those bytes, fixed producer identity trusted-research-retrieval@v1, and delegation to the generic kernel.

That adapter is intentionally not a retrieval implementation. The central reviewer still lacks a canonical, authenticated retrieval handoff that can supply those exact bytes without letting model prose, normalized text, snippets, citations, browser rendering, or mutable URLs impersonate producer evidence. Noema must not copy a central retrieval client or invent external research truth.

Owner boundary

This issue owns the shared reviewer/control-plane producer and its machine handoff. It must not move provider/model routing out of contextual-orchestrator, source/domain truth into Noema, or treat Zotero/OA/browser/search metadata as self-authenticating evidence. noema#555 remains the consumer/admission owner. Separate execution framing remains .github#2086.

Acceptance

Provide one versioned canonical trusted-research-retrieval producer/handoff that, before model review publication:

• captures exact retrieved response/content bytes with bounded size/time and explicit retrieval policy;
• records canonical source URI plus the immutable/content-addressed revision actually consumed;
• supplies exact non-empty excerpt bytes selected from those captured bytes, never a decoded/normalized/paraphrased substitute;
• binds repository, current head, workflow ref, run id/attempt and producer version to the out-of-band manifest path expected by Noema;
• makes redirects, content encoding/decompression, MIME/charset handling, partial/range responses, cache/revalidation and retry semantics explicit so the bytes being cited have one reproducible identity;
• fails closed on truncated/oversized/stalled responses, mutable-revision ambiguity, unsupported URI schemes, redirect/origin-policy violations, excerpt-not-present, duplicate/ambiguous handoff identity, and handoff substitution;
• preserves source licensing/access restrictions and does not upload private/controlled source bytes into a public artifact merely to satisfy review evidence;
• includes hostile fixtures for Unicode normalization differences, whitespace-only lookalikes, repeated excerpts, binary/invalid-UTF-8 content, compressed responses, redirects, stale cache/revalidation, oversized bodies and attacker-controlled citation text;
• keeps research evidence distinct from source-code evidence and sandboxed_verify execution evidence;
• publishes an immutable/versioned producer contract before Noema or central consumers expand the accepted producer version.

End-to-end completion evidence

After owner-side source is protected and versioned, noema#555 should consume that released producer contract through the existing trusted-research-retrieval@v1 adapter, produce an authenticated manifest on a real central-review path, and then bump only an immutable released Noema contract into the central consumer. The unchanged original external-evidence corpus must rerun RED→GREEN. Mutable sibling branch/PR consumption is not completion.

Related: ContextualWisdomLab/noema#555, Draft ContextualWisdomLab/noema#607, .github#2086.