Fix: Optimize token usage for scratchpad tool

[ea-rus]

Fixes ENG-626 — recall_scratchpad_wisdom() had no size cap and injected content unconditionally into the scratchpad tool description, which is resent on every LLM call. On a heavy real user this grew to ~103K chars, becoming the largest single component of the per-call payload and burying the tool's own safety-critical timeout/progress() contract.

Also fixes two related bugs found while investigating: a shared-singleton mutation bug that leaked memory content across sessions, and a duplication issue where the same procedural knowledge was billed twice per call (system prompt + tool description).

Changes

Static contract ordering (tool_defs.py)

• Moved the IMPORTANT timeout/progress() block right after the Actions list (where exec/cell are defined) and before the capabilities paragraphs — previously it sat right before the unbounded memory injection, so growth pushed it further from view on every turn.

Token budget for recall_scratchpad_wisdom() (hippocampus.py)

• "when" rules are now pulled via get_rules() (clean Engram objects) instead of scraping recall_rules()'s raw markdown, which was leaking <!-- confidence:... ts:... --> metadata comments verbatim into the injected text.
• Rules + lessons are sorted by confidence tier (high → medium/unset → low) then recency within a tier, and trimmed to token_budget (default 2000/scope, ~4 chars/token) via a shared _filter_by_token_budget() helper (also used by get_lessons()).
• Fixed a correctness bug: the method's own docstring says "relevant to scratchpad execution," but it pulled every when rule regardless of topic. Added the same "scratchpad" in text/topic filter already used for lessons (_is_scratchpad_related()).

Fixed SCRATCHPAD_TOOL singleton mutation (session.py)

• _build_core_tools() aliased the module-level SCRATCHPAD_TOOL singleton and mutated .description in place (scratchpad_tool = SCRATCHPAD_TOOL). In a process serving multiple sessions/users, each session's injected wisdom compounded on top of the previous one's, forever, until process restart — a real risk of leaking one user's memory into another user's tool description within the same worker.
• Fixed with scratchpad_tool = replace(SCRATCHPAD_TOOL) (shallow copy via dataclasses.replace).
• Added a regression test that builds tools for two sequential sessions and asserts the singleton's .description is untouched and no cross-session leakage occurs.

Removed duplication between system prompt and tool description (cortex.py, hippocampus.py)

• build_memory_context() (system prompt) and get_scratchpad_context() (tool description) were both pulling from the same get_rules()/get_lessons() data with no coordination — any scratchpad-related when rule or lesson was billed in both places on every call.
• Added exclude_scratchpad / exclude_scratchpad_when parameters to get_lessons() / get_rules() (default False, so recall_scratchpad_wisdom() and the existing recall_rules() caller in session.py — an ACC dedup check that needs the full rule set — are unaffected).
• recall_lessons() now always excludes scratchpad-related entries; build_memory_context() calls get_rules(exclude_scratchpad_when=True) / recall_lessons(token_budget=1000) directly, keeping the LLM-based rule-relevance filtering and episodic-memory read-logging paths intact.

Testing

• tests/test_hippocampus.py, tests/test_cortex.py, tests/test_chat_scratchpad.py — new coverage for: budget truncation, confidence/recency ordering, the scratchpad-relatedness filter on rules, singleton-mutation regression, and system-prompt/tool-description exclusion in both directions.
• Full suite: 1170 passed (one pre-existing e2e flake, test_large_input_no_crash, reproduces identically on main — unrelated to this change, confirmed via git stash).

Not in this PR

• Relocating long-tail lessons to recall_skill (pull-based) instead of push-injecting into the tool description — larger change, left as follow-up.
• Filtering user/task-specific trivia (e.g. campaign names) at write time in the consolidator/cerebellum — separate from the read-time budget/sort work done here.

Fixes: https://linear.app/mindsdb/issue/ENG-626/memory-consolidation-inflates-scratchpad-tool-description-without