feat(plugins): optional yantrik_memory — semantic memory + supersession (opt-in)

[spranab]

What this adds

An optional, off-by-default plugin (code_puppy/plugins/yantrik_memory/) that
adds a YantrikDB-backed semantic/learning memory layer alongside puppy_kennel
— not a replacement. Heads-up discussion: #464.

It mirrors the kennel's plugin contract (load_prompt, user_prompt_submit,
agent_run_end, register_tools, /yantrik command), and:

• distills durable facts/preferences from natural conversation (ignores chatter),
• supersedes outdated facts via correct() instead of accumulating them,
• surfaces query-relevant memories passively each turn,
• reinforces options that keep getting accepted.

Why (honest framing)

The kennel's P0 prefs band already handles basic fact recall well — so on "is the
fact recalled," a prefs-band baseline ties this; I'm not claiming better
recall. What an append-only prefs store can't do, and this adds:

• Supersession — after "we rebranded blue→green", the kennel's P0 keeps both
  (it's append-only), so the agent gets contradictory context; this drops the stale
  value. Verified stable from 1k→10k memories (the append-only band still carries
  both at 10k; this carries only green).
• Passive semantic recall — the kennel's passive layer is "all prefs + recent";
  finding the right past artifact otherwise relies on recency or the agent calling
  the BM25 search tool.

Safety / footprint

• Opt-in: default OFF; enable with /yantrik enable.
• Fail-soft: if yantrikdb isn't installed (or it's disabled), the plugin is
  inert — registers no callbacks, never crashes boot or a turn.
• Zero changes to existing code — a new plugin directory only.
• One optional dependency (yantrikdb); embedder is configurable down to a bundled
  no-ONNX option.

Testing

• Integration test (test_integration.py): distills facts, supersession works
  (recall's current band = green, stale blue gone), chatter adds nothing.
• Live load test: loads cleanly alongside all 35 builtin plugins (incl. the
  kennel); via the real callback bus, on_user_prompt_submit distills and
  on_load_prompt injects the correct (superseded) recall block.

Known limitations (full disclosure)

• The distiller runs a local-LLM call + a short flush per user turn — correct but
  not free; worth making async. Optional and model-configurable.
• Not yet run in a full interactive TUI session (validated via the real loader +
  callback dispatchers).
• YantrikDB's automatic conflict detection doesn't currently fire on free-text
  attribute-value updates, so supersession uses an explicit correct() at the app
  layer (the distiller detects the update). The engine side is being addressed
  separately.

Happy to adjust to your plugin/dependency conventions — flagging this as a starting
point for discussion as much as a merge candidate.