claw-wall: expose stable source handles in channel awareness

[mostlydev]

Summary

PR #262's digest-backed channel-memory design depends on stable source handles in the current Phase 1 channel-awareness surface. The raw awareness feed currently renders lines as timestamp + author + text, which is not enough provenance for a later digest block to round-trip to exact Discord messages when several authors post in the same minute or reports are chunked.

Implement the prerequisite as a tight Phase 1 follow-up before building the channel-memory adapter.

Motivation

Digest-backed awareness must be source-backed. A derived block can only be trustworthy if it points back to exact retained channel messages and the retrieval tools can fetch those exact messages. Timestamp-only references are ambiguous in busy Discord channels and make the Phase 2 digest design unsafe.

Scope

• Emit a stable source handle for each message in channel-awareness output.
• Use the existing Discord message ID when available; format it with the channel ID so the handle is unambiguous across channels.
• Keep channel-awareness as kind=raw_window; this issue does not implement raw_window+digest.
• Ensure search_channel_context and get_channel_messages expose the same stable handle in retrieval results or metadata, without removing the existing id / channel_id fields.
• Preserve current cursorized channel-context behavior unless the implementation chooses to share the same formatter safely.
• Update provider-visible examples/tests so future digest work can cite channel_id/message_id and retrieve exact source.

Constraints

• No new public pod YAML surface.
• No LLM work or channel-memory adapter in this issue.
• Channel ACLs remain enforced by the existing generated allowlist and claw-wall service token path.
• Keep the output readable for models; the source handle should be compact and mechanically parseable.
• Public docs/issues should stay generic and not reference named downstream deployments or agents.

Acceptance Criteria

• Unit tests show channel-awareness includes stable handles for raw messages.
• Retrieval tool tests show exact message lookup returns the same handle used by awareness provenance.
• Existing channel-context cursor tests keep passing.
• A spike test path verifies the claw-wall Docker image surfaces stable handles after Discord backfill. A fake Discord spike is the minimum; a real Discord pod spike is required before considering the Phase 2 channel-memory work proven.
• git diff --check and the relevant Go test packages pass.

Related

• PR Plan + ADR for channel-memory capability and digest-backed channel-awareness (issue #232 Phase 2) #262: channel-memory and digest-backed awareness design.
• Issue cllama: guarantee bounded 24h channel awareness independent of cursors and restarts #232: original bounded channel-awareness contract.
• Follow-up after this issue: examples/channel-memory with SQLite source retention and derived sparse blocks.

[mostlydev]

Deferred prerequisite note: #263 intentionally lands only stable source handles. The content_hash / edit-version identity called out in the #262 channel-memory plan is deferred to the channel-memory adapter phase, because today claw-wall keys Discord messages by channel/message id and does not observe edit revisions. The adapter/digest implementation should add the source_messages content-hash semantics when it introduces durable source rows.