✨ feat(user-memory): make ProfileExtractor target-aware in multi-speaker conversations

[LoveS0ph1e]

Summary

ProfileExtractor.aextract already receives a sender_id - the user the profile is for - but that id never reaches the extraction prompt. INIT and UPDATE are rendered with {conversation_text} alone, so the model is asked to "build the user's profile" from a transcript without being told which speaker is the subject. In any slice with more than one speaker - and a plain user <-> assistant exchange already qualifies - facts from another participant, or the assistant's own persona, can be attached to the profile owner.

The repository already anticipates this. TEAM_PROFILE_UPDATE_PROMPT carries a {target_user} block and explicit speaker-attribution rules, and its module docstring notes it is "not yet consumed by ProfileExtractor - kept for future minor extractor releases." This PR realises that intent on the active INIT and UPDATE paths, and generalises it: the problem is not specific to group chats - every conversation has at least a user and an assistant.

Changes

• _init_extract / _update_extract thread target_user=sender_id into render_prompt.
• The INIT and UPDATE prompts (en + zh) gain a **TARGET USER: user_id={target_user}** block and a sharper attribution rule - extract the target's facts only; every other participant, the assistant included, is context, never the subject. The wording mirrors the existing TEAM_PROFILE_UPDATE_PROMPT.
• _render_conversation is unchanged: every line is already tagged (user_id:...), which the prompt now references to attribute by speaker - the same signal TEAM_PROFILE_UPDATE_PROMPT relies on.
• COMPACT is left alone: it re-summarises an existing profile and never sees a conversation.

Backward compatibility

render_prompt fills placeholders by literal substitution, so the extra target_user= argument is a no-op for any template that does not reference {target_user} - including a caller-supplied prompt= override. Single-user extraction is unaffected.

Tests

• The 17 existing ProfileExtractor tests pass unchanged.
• 2 tests assert the target sender_id is rendered into the INIT and UPDATE prompts. As a discrimination check they fail against the unmodified extractor and pass with this change.

ruff check / ruff format --check are clean; the profile test suite is 19 passed against a clean everalgo-user-memory==0.3.1. mypy / pyright were not run locally (no workspace tooling) - please let CI confirm.

---

Rebase note (2026-07-02): This branch was rebased onto the anti-bloat HARD RULE PR (#3).
Rule 6 now carries the HARD RULE version with IF/THEN logic and WRONG/RIGHT examples, verified
in production (deepseek-v4-flash: 3/3 clean vs baseline 0/3). No other content changed.

[LoveS0ph1e]

Superseded by #4. The target-aware extraction core is preserved there, re-cut as a minimal diff (no whole-file rewrite), extended to zh parity, with a genuine multi-speaker test (asserting attribution keys on user_id, not display name) and an English code comment on the timestamp coarsening. Closing in favor of #4.