feat(viking): defer viking_remember to turn-boundary capture with memory-diff steer

[Leoyzen]

Summary

Redesigns viking_remember (OpenSpec change viking-remember-deferral) from a synchronous messages=[...] tool into a deferred, zero-argument capture:

• Tool call only queues the optional reason; the next before_model_request drains the real conversation (shared cursor with auto_ingest, disjoint ranges, remembered-capture runs before auto-ingest in the chain) into a remember-* session with <memory-intent> markers.
• Sanitization is unconditional (deferred capture runs after recall/profile injection).
• Commit is synchronous (HTTP round-trip only; Phase 2 extraction not awaited). On success: cursor advances + reasons clear. On failure: cursor and reasons are retained for retry (capped at 3 consecutive failures).
• A background task polls the extraction task and steers the memory diff (added/updated/deleted URIs) into the session via session_pool.steer_from_background_task (survives run boundaries), gated by new config remember_notify (default True).
• after_run now flushes pending remember intents plus the trailing [cursor, end] messages — closing the historical gap where the final assistant message was never ingested. The dead _pending_conversation field is removed (was written, never read).

Bugs fixed along the way

1. Dead _pending_conversation — final assistant turn was silently never ingested.
2. commit_session retention kwarg renamed keep_recent_turn_count → keep_recent_count (the SDK's real name; the old name raised TypeError against a live server whenever auto_ingest_keep_recent_turns > 0 — hidden by the canned mock).
3. source_type was a dead parameter — verified create_session/commit_session have no such SDK surface; provenance is now expressed via the session-id prefix (remember- / ingest-), per updated spec.

SDK discovery

openviking_sdk.AsyncHTTPClient signatures were inspected directly: create_session(session_id, telemetry, memory_policy), add_message(session_id, role, content, ...), commit_session(session_id, telemetry, *, keep_recent_count). No source_type surface exists.

Testing

• 434 viking tests pass (baseline 420) — unit + integration; new coverage: deferral, reason merge, retry semantics (cursor/reasons retention + cap), sanitize-always, provenance prefix, before_model_request chaining with auto_ingest disabled/enabled (disjoint ranges), after_run flush (final-assistant + last-moment remember + no-capture-when-all-disabled gate), steer positive/failure paths, task-poll failed-status.
• Ruff + mypy clean.
• Full suite: 6792 passed; only failure is the pre-existing test_providers/test_multimodal.py::test_vision (ALLOW_MODEL_REQUESTS gate) — reproduced on main.
• Deferred (need a live Viking server, human step): VCR cassettes (tasks 7.7/8.2–8.4). The notify pipeline is deterministically unit-tested; record with uv run pytest tests/capabilities/viking/ --record-mode=once and commit the cassettes when a server is available.