feat(user-memory): target-aware profile extraction + anti-bloat durable-abstraction rule (en+zh)

[LoveS0ph1e]

Summary

This PR makes ProfileExtractor target-aware in multi-speaker conversations and adds a durable-abstraction (anti-bloat) rule to the profile prompts, applied symmetrically to both English and Chinese. It consolidates and supersedes three earlier, overlapping PRs (#1, #2, #3) into one clean, minimal, reviewable change.

Motivation

Two independent quality problems surfaced when running ProfileExtractor on real group chats:

1. Cross-speaker contamination. In a multi-person transcript, facts stated by another participant — or by the AI assistant itself — could be absorbed into the wrong user's profile (including the assistant's own persona leaking in as a "user trait").
2. Profile bloat. Over many UPDATE cycles, description fields accumulated dated, event-log-style entries ("On 2026-06-29 …; on 2026-07-01 …") instead of stable generalizations, diluting the profile and hurting readability/retrieval.

What changed

• Target-aware extraction (profile.py, prompts/{en,zh}/profile.py): the sender_id already passed to aextract is now threaded into the INIT and UPDATE prompts as target_user. Each conversation line is tagged (user_id:...); the prompts instruct the model to attribute every fact to its stated speaker and to treat all other participants — and the assistant — as context, never the profile's subject. Attribution keys on user_id, not the display name.
• Anti-bloat "durable abstraction" HARD RULE (INIT + UPDATE, en + zh): each description must be a timeless generalization — no date/weekday/clock time (coarse time belongs in evidence); repeated instances of a pattern must be folded into one existing item via update rather than appended as new dated clauses. Includes WRONG/RIGHT examples and a pre-output self-check.
• Timestamp coarsening (_render_conversation): per-message timestamps are rendered at date granularity (YYYY-MM-DD). Profile extraction needs no intraday precision; this removes second-level noise, avoids UTC/local intraday misreads, and reinforces the no-clock-time rule.
• Test: a multi-speaker test asserts the target user_id is threaded into the prompt and that _render_conversation tags each speaker with their own user_id (so attribution keys on id, not name).
• CHANGELOG updated under [Unreleased].

Backward-compatible: render_prompt performs key-wise substitution and tolerates extra fields, so a caller-supplied prompt= override that omits {target_user} remains a no-op. No public signatures change.

Relationship to prior PRs (this supersedes #1, #2, #3)

These changes previously existed as three entangled PRs against the same two files; consolidating them removes mutual conflicts and cleans up the issues found in an audit:

• 🐛 fix(user-memory): strengthen anti-bloat rules with HARD RULE blocks and concrete examples #3 (anti-bloat HARD RULE) — its content is a byte-exact subset of ✨ feat(user-memory): synthesise a holistic Profile.summary instead of copying the first fact #1 and ✨ feat(user-memory): make ProfileExtractor target-aware in multi-speaker conversations #2 and is fully included here (now also in zh, with a test and CHANGELOG). Superseded.
• ✨ feat(user-memory): make ProfileExtractor target-aware in multi-speaker conversations #2 (target-aware extraction) — the sound core is preserved here, but re-cut as a minimal diff (no whole-file rewrite), extended to zh parity, with a real multi-speaker test (the original asserted only string substitution on a single-speaker fixture) and an English (not mixed-language) code comment on the timestamp change. Superseded.
• ✨ feat(user-memory): synthesise a holistic Profile.summary instead of copying the first fact #1 (holistic Profile.summary) — intentionally not carried forward. Its summary implementation had a blocking defect (an undefined llm_summary reference on the UPDATE path) and bundled the other two PRs' changes. The valid insight — that copying explicit_info[0] as the summary produces a misleading pseudo-summary — remains open; a clean upstream fix would add a summary field to the prompt output schema (so _build_summary's fallback is no longer the only source). Left as follow-up rather than shipped unproven. Superseded (closed in favor of this PR).

Validation

• Prompt modules load; {target_user} and the anti-bloat rule are present in INIT and UPDATE for both en and zh.
• render_prompt substitutes target_user, unescapes the literal JSON braces in the examples, and leaves no stray placeholders.
• The added multi-speaker test asserts attribution keys on user_id.
• The anti-bloat rule was validated in downstream production use (dated-description rate dropped substantially on a real bloated profile). Note: the repo's uv workspace test harness was not run in the authoring environment; CI should exercise the full suite.