fix(review): bind background_review origin for same-session injection turns

[kokhlo]

Fixes #107850

Root cause

When a review prompt ("Review the conversation above...") is injected into the foreground agent's loop without forking (v0.21 same-session injection path), the foreground agent._memory_write_origin stayed at the default "assistant_tool", causing is_background_review() to return False and bypassing the curator / skill-authoring gates that rely on it (user-owned-skills guard, read-before-write, skill ledger [auto] tag, approval staging).

Fix

turn_context.py:891 already binds agent._memory_write_origin onto the write-origin ContextVar at turn start. This PR adds detection for review-prompt injection: when the foreground agent receives a user message starting with "Review the conversation above", override the origin to "background_review" for the duration of that turn.

Both the fork path (build_cache_parity_fork already sets review_agent._memory_write_origin = "background_review") and the same-session injection path now share the same origin, so all guards fire correctly.

Evidence

Before: same-session review writes landed in ~/.hermes/skills/ without curator/gate review (issue evidence: 2026-09-10 18:08–18:12, platform=cli, skill_manage create passed the gate).

After: is_background_review() returns True for both paths; skill guards apply.

Test coverage

✓ 23 tests in test_turn_context.py (turn prologue, existing paths)
✓ 4 tests in test_skill_provenance.py (origin ContextVar)
✓ 10 tests in test_background_review_memory_scope.py (fork path, attended/unattended)
✓ 2 NEW tests in test_same_session_review_origin.py (injection detection logic)

39 tests passed — fork path, normal path, and injection path all verified.

[alt-glitch]

This was generated by AI during triage.

Related: #107850 (the issue), #52849 (merged: sets _memory_write_origin='background_review' on the curator fork), #61088 (pre-seeds the origin before AIAgent.__init__).

Note for reviewers: on current main every review path (_run_review_in_thread -> _run_review_fork -> build_cache_parity_fork, and agent/curator.py) runs in a fork that already carries _memory_write_origin='background_review'; the fork inherits agent.platform and shares the parent session_id, which matches the platform=cli / same-session-id log lines cited in the issue. A same-session non-fork review path was not found. Sniffing the Review the conversation above prefix in build_turn_context would also tag a real user's turn as a background review if they type that phrase.