Implements AACE RP 29R-03 §3.8 Modeled / Subtractive / Single
        Base method (paired with MIP 3.3 Windows for the dual-method
        gap report per SCL §11.5). Validates a forensic windows
        analysis (MIP 3.3) by independently computing the same
        project drift via subtractive removal of delays from the
        as-built schedule.

        For each delay event, the as-built duration of every
        ``affected_activity`` is shortened by ``impact_days`` working
        days on that activity's calendar (or removed entirely if
        ``removal_method="remove"``), then CPM re-runs and the
        resulting "but-for" finish date is compared to the as-built
        finish. Cumulative pass removes ALL events at once for a
        project-level but-for finish. Every finish shift is stated in
        working days on ``workday_calendar``; the calendar-day span of
        the same shift is kept beside it, and ``day_unit_basis`` names
        the unit of every bare ``*_days`` key in the result.

        Use this tool when opposing counsel demands a but-for analysis
        or you need a dual-method validation pairing §3.3 (windows) with
        §3.8 (collapsed-as-built). For prospective fragnet insertion
        (MIP 3.7), use ``time_impact_analysis_fragnet`` instead.

        Args:
            as_built_xer_path:    server-side post-impact XER (after delays incurred).
            as_built_xer_content: full text of post-impact XER (alternative for hosted/remote use).
                                  Supply EXACTLY ONE of path/content.
            delay_events: list of event dicts. Each must have
                ``event_id``, ``affected_activities`` (list of
                task_codes), and ``impact_days`` (number of working
                days on the affected activity's calendar). Optional:
                ``removal_method`` ('shorten'|'remove'),
                ``responsible_party``, ``name``, ``description``.
            output_dir: optional output dir for HTML/CSV (tempdir if "").
            project_name: optional override.
            removal_method: global default 'shorten' or 'remove'.
            contractor_filter: when True, exclude contractor-caused
                events from the cumulative pass (owner audit mode).

        Returns:
            {
              "as_built_finish": "YYYY-MM-DD",
              "per_event_results": [{event_id, but_for_finish,
                                      impact_workdays_collapsed,
                                      collapse_state,
                                      duration_removal_basis,
                                      finish_driver_after_removal, ...}, ...],
              # impact_workdays_collapsed is the event's finish shift in
              # working days on workday_calendar. Each event also carries
              # impact_days_collapsed: the same shift in calendar days.
              # duration_removal_basis discloses WHAT duration was removed
              # and on what basis; finish_driver_after_removal discloses
              # WHAT drives the but-for finish (incl. whether it is bound by
              # the data-date floor) so a reader sees WHY the finish did or
              # did not move across data dates.
              "cumulative_but_for_finish": "YYYY-MM-DD",
              "cumulative_impact_workdays": int,  # working days on workday_calendar
              "cumulative_impact_days": int,  # the same shift in calendar days
              "cumulative_collapse_state": str,  # a 'locked_by_actuals' zero is not "no impact"
              "workday_calendar": str,  # the calendar every working-day figure is stated on
              "day_unit_basis": dict,  # each bare *_days key: 'working_days' or 'calendar_days'
              "dual_method_gap": dict | None,  # working days only
              "output_files": {...},
              "warnings": [...],
              "method": "AACE 29R-03 §3.8 (Modeled/Subtractive/Single Simulation)"
            }