Fix Honcho auto-injection for shared gateway venues

[fadelguy]

Problem

Honcho session isolation was already working at the gateway/session-key level, but shared gateway venues (Discord threads/channels, Telegram groups, Slack channels, etc.) still received global peer recall during auto-injection. That meant a public/shared venue could pull in user cards, user representations, AI identity context, and dialectic supplements that belonged to the user globally rather than to the current venue.

In practice this caused cross-thread / cross-venue contamination:

• a trading thread could get school / Gmail / ops memory
• shared venues could trigger unwanted initiative because stale global memory was injected as if it were relevant
• session summaries stayed local, but the auto-injected peer context did not

Fix

This PR makes Honcho auto-injection venue-aware:

• non-DM gateway venues now auto-inject session summary only
• user representation / peer card / AI representation / AI card are no longer auto-injected there
• dialectic prewarm + dialectic prefetch are skipped for those venues
• Honcho context fetches now support summary_only=True, so we avoid fetching peer-global context when the policy does not allow it
• DM / CLI behavior stays rich and unchanged

Tests

Added regression coverage for:

• shared venues only auto-injecting session summaries
• DM venues keeping richer auto-injection
• shared venues skipping dialectic prefetch
• Honcho session manager summary_only=True fetching only the session summary and skipping peer fetches

After this change

Shared/public venues behave like venue-local conversations instead of semi-global memory sinks. Threads keep their own session continuity, but they stop auto-pulling unrelated user-global memory from other threads, channels, and workflows.

[teknium1]

Thanks for isolating the shared-venue recall problem. The premise still holds on current main: agent/agent_init.py:1366-1410 forwards chat_type, while Honcho still formats peer and AI context in plugins/memory/honcho/__init__.py:562-589, fetches it without venue filtering at :677-690, and runs dialectic at :695-768.

Problems

• plugins/memory/honcho/session.py:186 introduces a stale peer-ID resolver. Current main's canonical resolver (plugins/memory/honcho/session.py:331-361) additionally handles user_id_alt, configured aliases, runtime prefixes, and collision-safe generated IDs. Replacing it would regress the runtime identity mapping discussed as landed in #32863 / #30077.

Suggested changes

• During salvage, retain the summary-only and dialectic-skip policy but route stub/session peer resolution through current main's _resolve_user_peer_id(key).
• Adapt the policy to current main's asynchronous prefetch lifecycle (plugins/memory/honcho/__init__.py:653-690) and split the unrelated observation/status changes unless they have a separate verified need.

Automated hermes-sweeper review.

[teknium1]

Please reuse the canonical peer resolver when salvaging this helper. Current main's _resolve_user_peer_id() also honors user_id_alt, userPeerAliases, runtimePeerPrefix, and collision-safe generated IDs; this older branch-local logic would drop those identity guarantees.

[fadelguy]

Thanks — salvaged this against current main in 7485c75c.

The updated branch now:

• retains the canonical _resolve_user_peer_id() path and its runtime-ID alias/prefix/collision handling;
• threads summary_only=True through the current asynchronous context-prefetch lifecycle;
• skips dialectic prewarm/prefetch only for shared non-DM venues; and
• excludes the unrelated observation/status changes from this salvage.

Added focused regression coverage for summary-only retrieval and shared-venue vs DM formatting. tests/honcho_plugin/test_session.py passes (121 tests), along with ruff and compile checks.

[fadelguy]

Rebased and refreshed this PR against current main (head 38e9aed2c7).

The update preserves the shared-venue isolation policy through the current async prefetch lifecycle:

• shared gateway venues inject only their session summary;
• cached rich context is filtered before prompt formatting;
• peer-global dialectic prewarm/prefetch is disabled in shared venues;
• Honcho observation settings are scoped per session; and
• internal memory-file / identity tags are stripped from visible output.

Verification: 73 passed across the focused Honcho and context-scrubber tests; ruff, py_compile, and git diff --check pass.

Could a maintainer please re-review?