feat: add result metadata and structural list truncation to execute_code

[juliusguay]

Summary

Adds result_type, result_count, and JSON-aware list truncation to the
execute_code tool response. This establishes a result-metadata contract that
anchors the LLM's narrative and caps unbounded payloads at clean element
boundaries.

Problem

execute_code returned raw results with no payload cap. A 300-member force
extraction returned all 300 items unbounded. If a downstream character truncator
cut the JSON mid-list, the LLM would attempt to "complete" the malformed structure
with invented data — a fabrication pattern with real consequences in structural
analysis workflows.

Without a count field, there was no way for the LLM to distinguish a truncated
list from a complete one. A model reporting "38 members failed" from a truncated
result where only 38 were visible is indistinguishable from one that saw all members.

Changes

• MAX_RESULT_ITEMS = 200 added to sandbox/const.py
• sanitize_output in sandbox/stdio_helpers.py now truncates lists at element
  boundaries and appends a truncation notice as the final item (no mid-JSON cuts)
• _classify_result() runs before sanitize_output in executor.py so
  result_count reflects the true pre-truncation total
• result_type and result_count added to ExecutionResult and emitted in
  to_dict()
• execute_code docstring and server instructions updated to declare
  result_count as the authoritative count

Token impact

Scenario	Before	After	Savings
300-member force extraction, single turn	~4,500 tokens	~3,012 tokens	33%
3-turn session (history re-sent each turn)	~13,500 tokens	~9,036 tokens	~4,500 tokens

For large models where the list exceeds 200 items, result_count becomes the
only way to know the true total — the LLM can accurately report "5,000 members
analyzed, showing first 200" without fabricating counts.

For small models (< 200 items), no truncation triggers and result_count adds
~2 tokens of overhead only.

Tests

11 new tests in TestResultAnchoring covering:

• Type classification for all result types (scalar, bool, string, list, dict, null)
• COM tuple → "list" classification (COM APIs like GetPrimaryLoadCaseNumbers return tuples)
• Pre-truncation count invariant: result_count == true total when list is capped
• Truncation notice appended as final element (not mid-array)
• Error paths return result_type: "null"

All 132 executor tests pass.

---

Questions or feedback: julius.guay@bentley.com

[silvestre-perret-bentley]

Solved by #18.

Agreed to close this PR in an internal discussion with @JJGIV2010.