docs(retrieval): define feature retrieval contracts

10-prd/behavior/capabilities.md

@@ -18,7 +18,17 @@
 ## Capability: Semantic retrieval
 
 - related claim(s): expose info-base information for retrieval and downstream use; retrieve semantically related info-base entities
-- non-goal: semantic retrieval does not generate an answer, create transient information units, silently maintain its index while reading, or imply that feature and graph-navigation retrieval are implemented
+- non-goal: semantic retrieval does not generate an answer, create transient information units, silently maintain its index while reading, or subsume feature and graph-navigation retrieval
+
+## Capability: Lexical feature retrieval
+
+- related claim(s): expose info-base information for retrieval and downstream use; recall blocks from lexical feature evidence
+- non-goal: lexical feature retrieval does not generate an answer, recursively index the graph, replace semantic or graph-navigation retrieval, or imply perceptual/hybrid retrieval is implemented
+
+## Capability: Media interpretation
+
+- related claim(s): organize information in one reusable info-base; recall blocks from lexical feature evidence
+- non-goal: system-driven interpretation is one additive organization approach; it does not replace faithful Resolver materialization, write retrieval records directly, or define all media organization
 
 ## Capability: Focal-block rumination
 

10-prd/behavior/claims.md

@@ -52,7 +52,16 @@
 - evidence expectation: a real provider embeds graph state collected through ordinary Memos, RSS, Atom, storage, resolver, and organization boundaries; judged primary entities enter the global top three and outrank explicit distractors.
 - source rationale: `../_drivers/market-and-user-pressures.md`, `../_drivers/business-and-service-objectives.md`
 - realization pointers: `../../20-product-tdd/semantic-retrieval-and-peer-capabilities.md`, `../../20-product-tdd/claim-realization-matrix.md`
-- impact on existing claims: this realizes semantic retrieval only; feature retrieval, graph-navigation retrieval, answer generation, and Chat InKCre remain independent product capabilities.
+- impact on existing claims: this realizes semantic retrieval only; feature retrieval, graph-navigation retrieval, answer generation, and Chat InKCre remain independent product capabilities even when another capability is also implemented.
+
+## Claim: InKCre recalls blocks from lexical feature evidence
+
+- claim intent: a person, Agent, or downstream capability should be able to recover an existing block from exact clues remembered from its label, content, metadata, transcript, subtitle, OCR, or useful interpretation.
+- evaluation dimensions: one bounded query returns existing blocks with an explainable literal or term-match reason and a plain excerpt; contiguous Chinese fragments do not depend on a Chinese tokenizer; missing or stale derived support is maintained separately from the read.
+- evidence expectation: ordinary collected and organized graph state proves exact identifiers, Chinese fragments, metadata-only clues, document bodies, faithful image/audio/video text, and model-authored media interpretation through the same retrieval contract.
+- source rationale: `../_drivers/market-and-user-pressures.md`, `../_drivers/business-and-service-objectives.md`
+- realization pointers: `../../20-product-tdd/feature-retrieval-and-media-interpretation.md`, `../../20-product-tdd/claim-realization-matrix.md`
+- impact on existing claims: this realizes lexical feature retrieval only; semantic similarity, graph-navigation, perceptual matching, hybrid composition, answer generation, and Chat InKCre remain independent capabilities.
 
 ## Claim: InKCre accepts memo-like capture through familiar clients
 

10-prd/behavior/rules-and-invariants.md

@@ -36,6 +36,18 @@
 - violation impact: block semantics become harder to reason about and downstream use loses clarity.
 - linked claim(s): organize information in one reusable info-base; expose info-base information for retrieval and downstream use
 
+## Retrieval projections remain derived and independently maintained
+
+- rationale: lexical records and embeddings accelerate application queries but do not become information authority or a hidden read-time organization path.
+- violation impact: a query could mutate graph/index state unexpectedly, stale support could be mistaken for authored truth, or collection/organization ownership could leak into retrieval.
+- linked claim(s): expose info-base information for retrieval and downstream use; retrieve semantically related info-base entities; recall blocks from lexical feature evidence
+
+## Faithful media text and interpretation keep different meaning
+
+- rationale: transcript, subtitle, and OCR aim to preserve source evidence, while a model-authored description or summary adds an interpretation for later use.
+- violation impact: users and downstream capabilities could no longer tell source-derived information from organization-authored meaning, and Resolver materialization would silently acquire organization authority.
+- linked claim(s): organize information in one reusable info-base; recall blocks from lexical feature evidence
+
 ## Exact source evidence outranks heuristic duplicate reduction
 
 - rationale: stable native identity can justify reconciliation, while content fingerprints and time cutoffs provide weaker and less stable guarantees.

10-prd/behavior/workflows.md

@@ -48,6 +48,22 @@
 - exception flow: unavailable projections and stale records remain absent until explicit maintenance; a missing profile or eligible provider fails explicitly; uncertain post-dispatch outcomes are not replayed automatically
 - observable outcome: at least one useful existing block or relation is returned under a stated ranking contract without an answer-generation step
 
+## Workflow: Lexical Feature Retrieval
+
+- actor: a person, Agent, or downstream application capability
+- trigger: the actor submits one textual clue remembered from an information unit
+- normal flow: the capability compares the clue with fresh Block-local label/text projections and returns a small ranked set of existing blocks with plain excerpts and explicit match evidence
+- exception flow: unavailable projections and stale records remain absent until independent maintenance; the query does not repair records or recursively pull text through graph relations
+- observable outcome: an existing target block can be recovered from an exact identifier, fragment, metadata clue, document body, faithful media text, or prior interpretation without generating an answer
+
+## Workflow: System-Driven Media Interpretation
+
+- actor: organization capability running through a capable peer
+- trigger: a typed one-shot job starts, normally after a configured Cron occurrence
+- normal flow: the capability selects media blocks still missing an interpretation, routes each modality to its configured Agent, and adds useful interpretation blocks and relations through ordinary graph authority
+- exception flow: one unavailable modality, provider failure, or honest no-output result does not block useful independent candidates; completed graph effects remain and the next job derives its candidate set from current graph state
+- observable outcome: useful image, audio, or video meaning is added as an explicit interpretation graph that later application capabilities may retrieve
+
 ## Workflow: Focal-Block Rumination
 
 - actor: a person or application surface choosing one existing block

10-prd/glossary.md

@@ -96,6 +96,18 @@
 - user-visible or business lifecycle language: finding or navigating information and its context.
 - notes on ambiguity with framework terms: an application capability may use a sink, index, embedding, or resolver projection without transferring graph authority.
 
+## feature retrieval
+
+- canonical business meaning: retrieving information from observable features that do not require semantic-similarity ranking or an already-materialized graph path.
+- user-visible or business lifecycle language: finding something from a clue remembered about it.
+- notes on ambiguity with framework terms: lexical and future perceptual retrieval are feature-retrieval families; graph-navigation owns facts already expressed as graph structure, and hybrid retrieval composes rather than redefines the primitive capabilities.
+
+## lexical retrieval
+
+- canonical business meaning: feature retrieval over a block's bounded textual projection, including labels, literal fragments, terms, document text, faithful media text, and explicit interpretations.
+- user-visible or business lifecycle language: finding an information unit by a word, phrase, identifier, filename, transcript, subtitle, OCR clue, or description.
+- notes on ambiguity with framework terms: lexical retrieval returns existing blocks; it is not Chat, answer generation, recursive graph indexing, or a promise of language-specific tokenization.
+
 ## memo-like capture
 
 - canonical business meaning: low-friction collection of thoughts, surrounding events, and small pieces of information through a memo client or collector.

20-product-tdd/claim-realization-matrix.md

@@ -15,6 +15,7 @@ Map product claims to participating units and clarify each unit's technical resp
 | Reliably collect RSS and Atom feeds | RSS extension maps bounded RSS 2.0/Atom snapshots to feed, item, enclosure, full-text, and semantic-content graph state | `core-py` | Real-protocol HTTP doubles, PostgreSQL graph acceptance, optional live endpoint smoke, exact replay/update/state/enrichment cases; not a feed-reader product |
 | Preserve email communication records for later use | Mail extension maps protocol-neutral Mail facts to source, mailbox, canonical email, participant, body, MIME, flag, and reply/reference graph state; remote MIME bytes are materialized lazily | `core-py`; `client-web` as Job peer, info-base navigation host, and resolver consumer | Real Dovecot/IMAP plus PostgreSQL proves ordinary collection, bounded backfill, remote materialization, and built-browser rendering; not a Mail-only inbox or complete email agent |
 | Accept memo-like capture through familiar clients | Memos extension provides the first memo-family backend access mode | `core-py`; `client-web` for deployment-scoped extension configuration | Backend write/read, comments, attachments, hot credential replacement, hot enable/disable, and graph round-trip |
-| Organize information in one reusable info-base | Explicit focal-block rumination is the first admitted organization approach; an Agent may add an ordinary graph interpretation while preserving the focal graph | `core-py`; `client-web` as an explicit trigger surface | Real-provider Resolver/Agent/Tool/graph journey and meaningful no-write boundary; not a complete organization taxonomy, periodic organizer, replacement, merge, or linking realization |
-| Expose information for retrieval and downstream use | Semantic retrieval returns one bounded global ranking of existing Blocks/Relations and score metadata; feature and graph-navigation retrieval remain future | `core-py`; `client-web` and future Agent/application consumers | Pinned real Memos/RSS/Atom/HTML/storage corpus, real provider embedding and rumination, four top-three/distractor judgments, local and delegated Peer journeys; not answer generation or Chat InKCre |
+| Organize information in one reusable info-base | Explicit focal-block rumination and system-driven media interpretation are admitted organization approaches; Agents may add ordinary graph interpretations while preserving existing graph authority | `core-py`; `client-web` as an explicit rumination trigger and retrieval consumer | Real-provider Resolver/Agent/Tool/graph journeys, meaningful no-write boundary, and bounded independent media candidates; not a complete organization taxonomy, replacement, merge, or linking realization |
+| Expose information for retrieval and downstream use | Semantic retrieval ranks existing Blocks/Relations by meaning; lexical feature retrieval recalls existing Blocks from explainable textual clues; graph-navigation and hybrid composition remain independent | `core-py`; `client-web` and future Agent/application consumers | Pinned real multi-source/document/media corpus, real embedding and multimodal providers, local/delegated Peer journeys, and browser List navigation; not answer generation or Chat InKCre |
 | Retrieve semantically related info-base entities | Resolver/Relation projections feed profile-scoped derived embeddings; one domain facade executes locally or delegates the exact capability to an eligible Peer | `core-py`; `client-web` Peer consumer | Freshness/invalidation/maintenance checks, global Block/Relation ranking, real provider quality gate, exact-target and failover/outcome-unknown protocol cases |
+| Recall blocks from lexical feature evidence | Block-local Resolver labels/text feed one derived record per Block; one facade ranks exact/substring/term evidence locally or through an eligible Peer | `core-py`; `client-web` Peer consumer and InfoBaseListView host | Exact identifiers, Chinese fragments, Mail metadata, PDF body, real image/audio/video faithful text, Agent interpretation, freshness/Jobs, delegated Peer, and built-browser navigation |

20-product-tdd/cross-unit-contracts.md

@@ -46,3 +46,13 @@ Record durable data and behavior contracts that span more than one unit reposito
 - Peer discovery and invocation remain separate. The generic Peer layer sees exact capability
   identities and opaque protocol payloads; each business capability keeps its typed request,
   result, and local execution semantics.
+
+## Feature Retrieval And Media Interpretation Contract
+
+- Block-local lexical projection，derived-record maintenance，explainable ranked results，
+  multimodal AI content parts，faithful Resolver text materialization，system-driven media
+  interpretation and List-host navigation are owned by the
+  [Feature Retrieval And Media Interpretation Contract](feature-retrieval-and-media-interpretation.md).
+- Resolver materialization and Organization may add graph facts，but the lexical retrieval
+  owner alone maintains lexical records。Exact synchronous retrieval uses the generic Peer
+  protocol without giving PeerManager the query or result meaning。