fix: isolate effective mapping policies in shared helpers

[AGiorgetti]

Change

Fixes #21.

• Separate mapper-wide deterministic name allocation from per-entry helper planning. Sharing now includes effective options, nullable shapes, and execution context.
• Preserve complete helper graphs for every entry and isolate entry-sensitive declared/constructor resolution, including rejected constructor candidates.
• Forward the initiating method through tracked calls so shared recursive helpers report the correct entry without enlarging tracker state.
• Add 46 regression cases and synchronize helper-identity documentation and Unreleased/Fixes.

Verification

• Test-first matrix: 44 cases on unchanged develop produced 38 failures and six controls, zero skips. Four tracker-state/parameter-collision checks rejected an intermediate field-based prototype; final parameter forwarding preserves the baseline tracker layout.
• Final tree: Release solution build has zero warnings/errors; all 614 generator tests, five runtime tests, and one integration test pass with zero skips. All three source sample smoke runs pass.
• Final package consumers: primary-package and standalone abstractions consumers pass. Trimming is infrastructure-blocked by the cloud sandbox's Unix-domain task-host socket denial (MSB4216/SocketException 13).
• Preceding candidate package suite: 220 passed, two infrastructure failures, zero skips, including all 12 framework/language consumer combinations. Trimming and Native AOT failed at the same task-host boundary. These results are not claimed as a final-head full-suite pass.
• PowerShell-dependent local tests are unavailable because pwsh is not installed; the cloud also lacks clang. No assertions or package dependencies were weakened.
• Independent read-only review found no correctness blockers. CRLF-aware diff check and consumer-skill validation pass.
• Exact-head ordinary CI run 37782972610 passed for 4b4c472. Linux and Windows each passed 891 tests with zero failures/skips, including trimming and Native AOT. All six Roslyn 4.8/4.14/5.9 matrix jobs passed 614 tests each with zero failures/skips. Linux sample smoke checks and all five benchmark smoke cases passed.
• Pack & Validate, Verify Release Artifacts, and Publish NuGet Packages were skipped in that ordinary PR run.
• The published tree was verified byte-for-byte against the tested local tree d88d3655d6c9139bae7af59f42a5b64d5275cf48. The PR has no merge conflicts and remains draft for human review.

Retained as draft for human review. No merge, release, tag push, package publication, or workflow dispatch was performed. Separate #22/#25/#26 fixes remain out of scope.

Prepared with OpenAI Codex.