claw-wall: push channel messages into channel-memory

[mostlydev]

Motivation

The channel-memory adapter should not pull from Discord independently. claw-wall already owns Discord channel surfaces, raw message identity, ACL checks, backfill/polling, and feed serving, so Phase 2 should use claw-wall as the source owner and push retained messages into channel-memory.

Depends on #265.

Scope

• Add optional compile-time/runtime wiring for claw-wall to discover a channel-memory ingest endpoint when the adapter is present.
• Push retained Discord messages from claw-wall to channel-memory after backfill and forward polling.
• Send stable channel/message identity, author metadata, timestamps, edit/delete state, content, and content_hash.
• Make push ingest idempotent and retryable without blocking raw-window channel-awareness serving.
• Preserve channel allowlist/visibility boundaries; channel-memory must not become a bypass around claw-wall policy.
• Add a repair/backfill path that can replay retained raw messages through the same ingest contract.

Out of scope

• Implementing the SQLite adapter itself beyond whatever test fake is needed.
• Serving digest-backed channel awareness.
• Direct Discord gateway ingest inside channel-memory.
• Release pin, changelog, or nav changes.

Acceptance

• Unit tests prove claw-wall emits the expected ingest payload for backfilled and newly polled messages.
• Failure tests prove channel-awareness remains fail-open to the raw window when channel-memory ingest is unavailable.
• Authorization/allowlist tests prove pushed data is scoped to the channels claw-wall is allowed to serve.
• A Docker/fake-service spike proves messages cross the claw-wall to channel-memory boundary.
• Public docs and issue/PR text stay deployment-agnostic.

Refs #262
Refs #263

[mostlydev]

Implementation note from the channel-memory adapter review: observed_seq in the adapter is arrival order, not source-platform edit chronology. The push-ingest wiring should preserve delivery order where possible and define behavior for out-of-order edit/delete events before relying on observed_seq as freshness.

[mostlydev]

Implementation review follow-up before merge: this slice only gives channel-memory first-sighting source-message ingest from claw-wall. observed_seq remains adapter arrival order, and claw-wall currently deduplicates retained messages by Discord message ID; it does not emit live edit supersession rows or delete tombstones from this source. Backfill captures already-edited content as seen at cold start, but #267/#268 must not assume live edit/delete signals are flowing until a gateway edit/delete source exists.

This is not a blocker for #266; it records the freshness boundary for the digest-serving and worker slices.

Refs #262 #265 #267 #268