fix: surface participant description in the passive roster (INT-943)

[AlexanderZ-Band]

Summary

• Surfaces each participant's description in the always-injected passive roster (build_participants_message), so the LLM can route by role without a roster tool call.
• Closes the gaps a review of this branch surfaced: add_participant()'s tool call and ExecutionContext's participant cache were both silently dropping/losing description on certain paths (see commit body for detail).
• Hardens the roster line against a self-authored description injecting fake extra roster entries or spoofing the trailing IMPORTANT: instruction line.

Test plan

• uv run pytest tests/ --ignore=tests/integration/ --ignore=tests/e2e/ -v — 4365 passed, 85 skipped
• uv run ruff check . && uv run ruff format --check .
• uv run pyrefly check
• Live E2E baseline smoke (test_reports_peer_description_from_passive_roster) — not run in this session; recommend running before merge to confirm the live backend's list_agent_chat_participants response actually populates description for agent participants (the Fern ChatParticipant type doesn't declare that field but has extra="allow", so it round-trips if the wire payload sends it — this is the one fact that needs live confirmation, not a code fix)

Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01Ldhhg7ykaMGPDEM9hzWppS

[linear-code]

INT-943

[amit-gazal-band]

Two things to address before merge: peer-authored descriptions land verbatim in every co-participant's system prompt, and the sparse-source merge guard doesn't match how the consumer defines "absent". Details inline.

[amit-gazal-band]

Follow-up on the two fixes in 5938298 — the merge change is fully resolved, the quoting has one small gap left. Detail inline.

Verified locally at this SHA: tests/runtime/ 687 passed. The unrelated failures in a full-suite run (test_logging_config.py, desktop_app) reproduce on the branch base / pass in isolation respectively — not this PR.

[amit-gazal-band]

Approving — both blocking findings are resolved, plus the follow-up nit.

• Merge absence check (5938298): snapshot.get(name) or existing.get(name) now matches the consumer exactly, with a regression test for the "" case.
• Description injection (5938298 + c6fff38): quoted value, a gated non-authoritative caveat, and quotes collapsed in the sanitizer before the length cap so the value can't close its own quoting. test_description_cannot_close_its_own_quoting asserts count('"') == 2, which holds for any future breakout rather than just the one example.

Verified at c6fff38: tests/runtime/ 688 passed, and I probed the sanitizer directly with quote-breakout / trailing-quote / all-quotes / newline+quote descriptions — exactly one quoted span in each.

One thing still open, not blocking and not a code defect: whether list_agent_chat_participants actually returns description on the wire. Fern's ChatParticipant doesn't declare it, so it only survives via extra="allow" if the backend happens to send it — and that endpoint is the only hydration source most rooms use. If it's absent, the roster renders description-less in the common path and the new E2E smoke fails, with no lookup_peers enrichment fallback to compensate. Worth confirming live before this ships rather than after.

Also noting for a follow-up (all out of scope for this PR): ParticipantTracker was removed from band.__all__ with no deprecation shim; the tool_result_hook plumbing in claude_sdk/tools.py and mcp/backends.py now has zero callers; and PASSIVE_ROSTER_DESCRIPTIONS_PROBE's "without calling any tools" may flake on tool-first adapters where band_send_message is the reply path.