This is the headline forensic tool — it computes per-window
        completion shifts, per-window slip registers (per-activity slip
        with a critical flag (total float of zero or less, or the driving
        path in a schedule with negative float)), per-window duration
        growth on critical-path activities, per-window per-party attribution
        (Owner / Contractor / Concurrent / Force Majeure / Unattributed),
        and cumulative project drift from baseline. The attribution math
        satisfies the CPP conservation check, per the AACE 29R-03
        §3.3.E.13 requirement that the summed per-period net impacts
        equal the difference between the first schedule update and the
        last schedule update used in the evaluation (per-party day
        buckets sum to project drift within ±1 day, no cascade-double-
        counting).

        Use this tool for the full multi-window forensic claim. If you
        already have a windows result and only want the per-window ×
        per-party grid view, call ``concurrent_delay_matrix`` instead.

        Args:
            schedules: list of dicts in chronological order. Minimum 2
                entries (baseline + at least one update). Each dict
                must contain ``label`` (str) and EXACTLY ONE of:
                  - ``xer_path``    — server-side filesystem path, OR
                  - ``xer_content`` — full XER text content.
                Use ``xer_content`` when calling a hosted MCP server
                from a remote client whose XER lives locally.
            project_name: optional override; auto-picked from XER if "".
            baseline_idx: which entry in ``schedules`` is the contract
                baseline (default 0 = first one).
            entitlement_milestone: optional task_code (e.g.
                "Ready for Takeover") — recorded on the result, not used
                for math.
            company: optional, the firm the dashboard and the DOCX report carry (title
                block, print cover). Blank unless supplied, and blank
                carries no firm: a hosted run is the caller's document.
            output_dir: optional dir for HTML dashboard / DOCX report.
                If "", a tempdir is used and dropped after — the
                dashboard / report paths in the response will point to
                the temp location (caller responsible for moving them).

        Returns:
            {
              "analysis": full dict from run_windows() with keys:
                "windows", "cumulative", "baseline_label", "data_dates",
                "attribution_summary", "mcpm_attribution", ...,
              "dashboard": path to HTML dashboard (server-side),
              "report":    path to DOCX executive report (server-side),
              "baseline_stability": {"worst_severity", "has_block", ...},
              "data_date_corrections": the record of each update whose
                  data date was moved to its latest actual, and of a
                  baseline with actual dates after its data date
                  ([] on clean files); each is also a data-quality
                  warning
              "analysis"."criticality_basis": the schedules read by
                  their driving path because they have negative float,
                  and those beside them ([] when no schedule has
                  negative float); the box they write is one
                  data-quality warning
            }

            On failure: {"error": "..."} with no schedules processed.