fix(hyperlight-codeact): pin version >=0.3.0 and handle Python 3.14+

Sandbox fixes (both monty + hyperlight):
- glob() returns workspace-relative paths (was absolute — caused retries)
- glob() supports brace expansion: {db,services}/**/*.py
- web_fetch() strips HTML to text via html.parser, caps at 20K chars
- uv run --quiet suppresses spinner that polluted tool results
- mcp_call always listed in tool reference (was missing at install time)

Skill documentation:
- Return types documented (glob→list, view→string, bash→dict)
- No os.path/os.walk — explicit prohibition
- One program, one bash call rule
- Backend-specific MCP syntax (monty: mcp_call(), hyperlight: call_tool)

Perf test infrastructure:
- Input token tracking from copilot process logs
- Context bloat measurement (tool_result_bytes)
- Cost estimation at GPT-5.4 pricing
- Failed/timed-out runs marked SKIPPED (not fake 100% reduction)
- Partial result capture (tool.execution_partial_result events)
- compare_results.py updated with context + cost metrics

13 benchmark prompts covering:
- Cross-file analysis (config audit, test coverage, import maps)
- MCP context bloat (same task with 0 vs 4 MCP servers)
- Real-world tasks (docstring coverage, function index, env var audit)
- Difficult cases that track where codeact still needs improvement

README updated with benchmark results table.
Unit tests for tool allow/deny configuration (14 tests).

Critical fixes:
- Path traversal: use os.sep in _check_workspace (prevents /workspace_evil bypass)
- Env var leak: MCP bridge now uses allowlist (PATH,HOME,LANG,etc) not full os.environ
- SSRF: web_fetch blocks non-http/https URLs; MCP bridge validates URL scheme
- File size: _view() rejects files >50MB before reading
- MCP response cap: _read_capped() limits to 10MB
- SQLite leak: connections closed and dict cleared at start of each run
- Silent exceptions: MCP config parse errors now logged to stderr

High fixes:
- curl: added --max-time 30 --max-redirs 5
- web_fetch truncation: [truncated] marker already present (verified)

Low fixes:
- Removed unused itertools import in monty _glob
- Fixed monty SKILL.md brace expansion docs (said fails, actually works)
- Fixed hyperlight web_fetch description and added max_length param

- SQL fetchall capped at 10K rows (both backends)
- Template injection: base64-encode values before passing to Python heredoc
- Process log extraction: retry loop + 20MB read cap
- Glob brace expansion: already correctly capped at 200 after merge (verified)

- Add --raw flag to both monty and hyperlight codeact scripts for
  direct stdout/stderr output instead of JSON envelope
- Add glob auto-exclude for common dirs (.venv, node_modules,
  __pycache__, .git, target, dist, build) with exclude_dirs param
- Raise glob cap from 200 to 10000 with truncation warning
- Add output minimization guidance to instructions template
- Add glob auto-exclude documentation to instructions

Add --raw output mode, glob auto-excludes, and output guidance