feat(core): loop suite, observability and memory P1 modules (split from #181)

[raymondginger2018-sudo]

Summary

The wiring part of #181, split per maintainer review so the two independent modules (keyring, classifier) can merge first. Contains the P1 course modules that touch existing behavior, kept together because they share the compaction/observability path:

• P1-5 compaction-as-memory: runner.py gains AgentRunSpec.compaction_summary_sink + _notify_compaction_summary; session.py builds the sink that appends summaries to the memory vault (memory_distill.py) and fires the memory.compaction.deposited event via observability/events.py.
• P1-4 tool-loop temperature: _DEFAULT_TOOL_LOOP_TEMPERATURE = 0.1 used only when the execution profile does not set one.
• P1-3 memory data boundary: memory.py wraps memory notes in a framed, explicitly-untrusted data block (injection_regression.render_data_block); the upstream framing assertion test_every_injected_instruction_source_is_framed stays green.
• P1-6/7/8 loop suite: loop/goal_file.py, optimizer.py, evaluation.py, retrieval_evaluation.py, cerebellum_optimizer.py, injection_regression.py, memory_retrieval.py.
• P1-2/9 MCP: server allowlist (mcp/naming.py + runtime.py) and tool description quality (tools/base.py + mcp/tools.py).
• agent_setup.py: system-prompt integrity clause.

Tests

• 132 module tests pass locally (observability, compaction-memory, memory_distill ×2, optimizer, evaluation, retrieval_evaluation, cerebellum_optimizer, goal_file, injection_regression, memory_retrieval, mcp allowlist, tool description quality)
• 17 memory regression tests pass, incl. test_every_injected_instruction_source_is_framed
• 78 harness/compaction regression tests pass (1 pre-existing env failure on clean base)

Dependency note

session.py imports core.observability.events (this PR); it already contains an explicit no-op fallback so merge order vs. #183 is irrelevant.

[raymondginger2018-sudo]

CI status note for the wiring split of #181:

• Python CI / Desktop CI / Linting: green (132 module tests + 17 memory regression + 78 harness/compaction regression locally).
• Security CI: the repo-wide pip advisory failure (PYSEC-2026-3721 on locked pip==26.1.2) — fixed in fix(security): bump sidecar pip 26.1.2 → 26.2 (PYSEC-2026-3721) #191, unrelated to this PR.

As noted in the description, this PR keeps the explicit no-op fallback for the core.observability.events import, so merge order vs #183/#188/#189 remains safe. @Zongwei9888 happy to take this one last, after the two independent modules.

[raymondginger2018-sudo]

Hi Zongwei, this is the loop suite + observability + memory P1 modules split from #181. Wires into existing behavior but keeps an explicit no-op fallback for the observability import so it is safe regardless of merge order. Rebased onto latest main (e0767d0), all tests green (3.12/3.13/3.14). The lint failure is a pre-existing N999 issue on upstream main, not caused by this PR. Ready for review. Thanks!