Add bounded function-group workflows and portable validation

[SamG-Coder]

Reconstruction work that spans unrelated classes or unnamed functions currently has to be organized outside ReAgent. This adds bounded function manifests that can be planned, executed, inspected, and tracked through the existing backend, selection, session, and validation pipeline. Build/test/runtime gates also gain argument-array execution for native Windows and POSIX.

Changes

• Add plan to collect seed functions and direct callees within explicit depth and function limits, without model calls. Preserve returned evidence and record incomplete exploration.
• Add reverse --manifest using the existing batch runner, including dependency ordering, retry limits, cumulative scratch validation, and cross-class resume.
• Preserve structured evidence gaps and full context records in the existing knowledge graph.
• Add evidence --manifest for linked per-function JSON packets and searchable TSV indexes.
• Add status --manifest for inventory coverage, stale-input reporting, external dependencies, and separate validation-check outcomes.
• Support shell-free argument arrays for build/test/runtime gates while preserving legacy POSIX shell strings. Extend doctor and add Windows Python 3.12 CI coverage.
• Fix Clang character offsets for CRLF source files and validate external evidence records more strictly.

These are general-purpose tool upgrades. No game-specific addresses, formats, assets, runtime code, or external project dependencies are introduced.

Validation

• Full local Windows/Python 3.12 suite: 169 passed, 9 skipped.
• Skips: five POSIX shell cases, three tests requiring Windows symlink privileges, and one POSIX process-group case.
• ruff check src tests examples passed.
• mypy --platform linux src/re_agent passed, matching the existing Linux CI target.
• Source archive and wheel built successfully.
• Installed-wheel smoke checks passed for version, planning, manifest dry-run, evidence export, status, and packaged prompt resources.
• Real C++ regression tests exercise compilation repair, differential failures, and cumulative cross-class restoration without modifying original source.

Live Ghidra and live model reconstruction were not exercised. Manifest coverage describes the selected inventory and configured acceptance policy; it does not establish whole-program completeness or semantic equivalence. External backend state without local exports cannot be fully fingerprinted, as documented.