feat(aidd-context): let learn ingest host-native memory as a source

Start at the aidd-context:10-learn source-selection and gather workflow, then compare the existing conversation, file, diff, and review sources. Read #455 and the Claude/Codex validation context to understand the host-agnostic boundary. Done means memory is an explicit source with confirmation and destination routing preserved, explicit failure when unavailable, unchanged existing sources, and validation on two supported hosts.

Problem

aidd-context:10-learn can currently learn from conversation, file, diff, and review sources.

AI coding hosts may maintain durable project knowledge across sessions. For example, a host may retain project facts, decisions, corrections, constraints, or session learnings in its own persistent memory.

That knowledge can contain exactly the kind of durable learning that 10-learn is designed to assess and promote into project memory, decisions, rules, contracts, or skills.

Today, once that knowledge only exists in the host's persistent memory and is no longer present in the current conversation, 10-learn has no source kind for consuming it.

This also creates a portability concern: host-native memory mechanisms differ between tools, while 10-learn is intentionally tool-agnostic

Scope

• Add host-native persistent memory as a supported source kind for aidd-context:10-learn.
• Keep the source semantic generic (memory), without exposing a Kilo-, Codex-, Claude-, or other host-specific source kind to 10-learn.
• Resolve the source through the current host's available persistent-memory mechanism when one exists and is readable.
• Feed the retrieved content into the existing gather → assess → write → sync workflow.
• Fail explicitly when the current host exposes no readable persistent memory; never silently substitute another source.
• Preserve the existing confirmation and destination-routing behavior: retrieved memories are candidates, not automatically promoted project knowledge.

Acceptance criteria

• 10-learn accepts memory as a source in addition to its existing source kinds.
• On a host with readable persistent project memory, 10-learn memory can gather durable learnings from that memory and pass them through the existing assessment flow.
• The 10-learn workflow does not contain a host-specific source kind such as kilo-memory, codex-memory, or claude-memory.
• Retrieved host memories are assessed with the same scoring, reconciliation, destination, and user-confirmation rules as other learn sources.
• A memory candidate can be routed to the existing destinations, including project memory, decision/ADR, contract, rule, or skill as applicable.
• No host memory is automatically promoted or written into AIDD artifacts without the existing confirmation step.
• If the active host has no readable persistent-memory mechanism, the source fails explicitly rather than falling back to conversation or another source.
• Existing conversation, file, diff, and review source behavior is unchanged.
• The implementation is validated on at least two supported hosts that expose persistent memory, so the source contract does not depend on one host's storage format.

Prior art in this repo

• aidd-context:10-learn already separates source selection from gathering and supports conversation, file, diff, and review.
• #278 introduced the requirement for 10-learn to consume material outside the current conversation.
• #455 implemented that direction by splitting source from gather and validating the workflow across Claude and Codex.
• #282 deliberately keeps 10-learn tool-agnostic: host/tool-specific placement of generated rules and skills belongs to their generators rather than to 10-learn.
• aidd-context:02-project-memory already separates canonical AIDD project memory from the wiring used to expose it to different AI tools.

This proposal applies the same separation in the opposite direction: host-native memory is an input source; AIDD project memory, decisions, rules, contracts, and skills remain the controlled destinations.

Out of scope

• Defining a new canonical AIDD memory format.
• Replacing aidd-context:02-project-memory.
• Synchronizing AIDD project memory back into every host's private memory store.
• Automatically scanning host memory on every prompt or session.
• Automatically converting host memories into rules, ADRs, contracts, or skills without user confirmation.
• Making 10-learn understand the internal storage format of individual hosts.
• Requiring every supported AI tool to provide persistent memory; hosts without readable memory may report the source as unavailable.