[Aschenbath]

After a builder writes partial output and returns an error, scons -i currently records that output as successfully built. The next invocation reports it as up to date; with --cache-force, the failed output also enters the shared cache.

Finish ignored failures using the Taskmaster path that updates scheduling state without calling the successful-build callbacks. Dependent tasks and shared side effects can continue, while the failed action does not save a successful signature or force its output into CacheDir.

The new end-to-end regression runs repeated real builds with -i, --ignore-errors --cache-force, and -i -j 2 --cache-force. It checks retries with outputs present and removed, multiple outputs, dependent and independent tasks, successful cache reuse, and failure after a previously successful build. The manual and release notes describe the behavior. Partial files are retained; this does not add the delete-on-error feature discussed in #4747.

Validation on Windows / Python 3.12.12:

• The new regression fails against the original source because a subsequent build returns success without retrying the failed action; it passes after the fix.
• Taskmaster, filesystem, CacheDir, CLI option and interactive tests: 129 scripts passed, 1 platform skip, 0 failed (130 selected).
• All 215 documentation XML files validate after running bin/docs-update-generated.py. Generated files are excluded from the change as requested by the repository guide.
• The new test passes Ruff lint/format checks; git diff --check passes. Existing diagnostics are unchanged from an independent baseline checkout: 54 Ruff findings in Main.py, and 1,353 mypy errors across the repository.

Contributor checklist:

• Added regression coverage for the changed behavior.
• Updated CHANGES.txt and RELEASE.txt; read README.rst.
• Updated the relevant documentation.

Developed and tested with substantial AI assistance (OpenAI Codex), disclosed in the commit trailer as well.

Fixes #1332.