fix(channel): filter hermes progress messages, deliver streamed final updates

[madtank]

Summary

Restores two ax-channel bridge behaviors that regressed when the TypeScript bridge was ported to Python axctl channel.

Root cause

1. The Python bridge subscribed only to message and mention SSE events; message_updated was dropped. Hermes-runtime sentinels seed a placeholder message ("Working…") on reply start and overwrite the same message_id in place as the real reply streams in. Without message_updated, the bridge saw the placeholder creation and nothing else.
2. The Python bridge had no filter for progress/chunk payloads, so every "Working…" / "Received" / streaming placeholder woke the Claude Code session.

Net effect: sessions got constant chatter from sentinels but never saw the actual replies without manually fetching via GET /api/v1/messages/{id}.

Fix

• Subscribe to message_updated in addition to message / mention.
• Allow message_updated to bypass seen-dedup only when the id has NOT been delivered yet (so final streaming updates supersede the placeholder but can't re-wake after delivery).
• Skip payloads where metadata.streaming_reply.final is explicitly false (structured runtime signal).
• Defensive regex fallback anchored with \Z: drop payloads whose first line (after stripping any leading @mention) is entirely progress-only — Working… / Received / Thinking... / Processing… with only trailing whitespace / dots / ellipses, or No response after <N>[smh] .... Prompts that merely start with those words like "Working-state cleanup proposal" are NOT dropped.
• Skipped payloads are NOT added to seen_ids so the subsequent final update for the same id still delivers.
• Also fixes the leading-mention strip regex to handle hyphenated agent handles.

Tests

Seven targeted tests in tests/test_channel.py:

• test_channel_skips_streaming_reply_non_final — structured metadata.streaming_reply.final=false filter.
• test_channel_skips_working_progress_message — defensive regex catches Working… / Received / Thinking... / No response after 5m ....
• test_channel_delivers_prompts_that_merely_start_with_progress_words — "Working-state cleanup proposal" still lands (regression for over-broad regex caught in review).
• test_channel_delivers_processing_webhook_errors_prompt
• test_channel_delivers_thinking_through_issue_prompt
• test_channel_delivers_message_updated_final — placeholder → final update round-trip delivers the real content.
• test_channel_skips_message_updated_for_already_delivered — duplicate finals don't re-wake.

All 20 channel tests pass. ruff check and ruff format --check clean. Full test suite: 224 passed.

Test plan

• CI green
• Install from branch, restart a Claude Code session, verify sentinel replies deliver without progress noise
• Regression check: a normal non-streaming message still delivers immediately

[madtank]

Code review note: one issue before merge. The PR did come from Orion's branch (orion/channel-streaming-reply-filter); the GitHub PR author shows as madtank, but the main commit is authored by orion and co-authored by Claude.

Blocking issue: ax_cli/commands/channel.py around the new progress-message regex skips any first line that merely starts with Working, Received, Thinking, or Processing. Because the pattern is not anchored to the end of the progress marker, a legitimate prompt like @anvil Working-state cleanup proposal, @anvil Processing webhook errors, or @anvil Thinking through this API issue would be silently dropped and not delivered to the channel.

Please tighten the fallback regex so it only suppresses known progress-only lines, e.g. exact/punctuation-only Working..., Received, Thinking..., Processing..., and No response after ..., while allowing ordinary user/agent messages that start with those words. Add a regression test for a real prompt such as @anvil Working-state cleanup proposal being delivered.

Everything else looks sound: subscribing to message_updated, not marking skipped progress chunks as seen, and delivering the final streamed update are the right fixes. CI is green after the format cleanup.