Fix model selection, cleanup dead schema, and rework Assist agent routing

[techartdev]

Summary

This PR bundles the low-risk fixes that are clearly ready plus a conservative rework of the useful parts of earlier proposals.

Included fixes

• fix the Active Model select so its value is actually passed to chat-card and Assist requests
• remove the dead attachments field from the openclaw.send_message schema/docs/translations
• align chat-card cache-busting URLs with the current integration version
• fix manifest documentation / issue tracker links so they point to the integration repo
• move event bus subscriptions into async_added_to_hass for cleaner HA entity lifecycle behavior
• extract duplicated recursive response-text parsing into a shared helper
• add a warning when the API client has to fall back to creating its own aiohttp session
• conservatively namespace Assist conversation IDs by agent ID to avoid cross-agent context bleed without introducing sticky persistent sessions

Issue / PR mapping

Fixes or addresses

• fixes Manifest links to your HA OpenClaw App (add-on) #15 by correcting manifest links
• fixes the broken model selector behavior identified in PR Fix model selector, dead code, event lifecycle, and cache-busting #14
• fixes the dead attachments service field identified in PR Fix model selector, dead code, event lifecycle, and cache-busting #14
• fixes cache-busting/version drift identified in PR Fix model selector, dead code, event lifecycle, and cache-busting #14
• fixes the event entity lifecycle concern identified in PR Fix model selector, dead code, event lifecycle, and cache-busting #14
• addresses Agent Id not handled #8 by improving Assist agent routing and preventing different agents from sharing the same conversation ID namespace
• documents the behavior behind Context settings #16 and clarifies context/model/agent behavior in the README

Explicit rework (not merged verbatim)

• reworks the useful part of PR feat(assist): sticky sessions and agent routing #10 into a smaller, lower-risk change:
  • keeps agent-aware Assist session namespacing
  • does not introduce sticky persisted sessions
  • does not override HA conversation flow more than necessary
• takes the useful warning/logging idea from PR Fix model selector, dead code, event lifecycle, and cache-busting #14 / PR Feature/debug logging #9 only where it is safe and directly actionable

Not included here

• PR feat: enable continue_conversation for Voice PE follow-up dialog #11 and PR Set x-openclaw-message-channel header for voice requests #13 were already merged into main before this branch
• the broader sticky-session architecture from PR feat(assist): sticky sessions and agent routing #10 is intentionally not included as-is
• the larger room-aware voice-context work from PR feat: inject device area context for room-aware voice responses #12 is intentionally not included here

Why this shape

The goal was to land the clearly-correct fixes now and only rework the risky bits where there was a clean, low-blast-radius version worth keeping.

Manual test checklist

• Select a model via the Active Model entity and verify chat-card requests use it
• Select a model and verify Assist requests use it
• Confirm different Assist agent IDs no longer share the same conversation ID namespace
• Confirm openclaw.send_message no longer advertises unsupported attachments
• Confirm event entities still fire correctly in the automation UI
• Confirm the chat card reloads after update without stale cached JS